[PeepholeOptimizer] Refactor optimizeUncoalescable logic

Reapply r242294.

- Create a new CopyRewriter for Uncoalescable copy-like instructions
- Change the ValueTracker to return a ValueTrackerResult

This makes optimizeUncoalescable looks more like optimizeCoalescable and
use the CopyRewritter infrastructure.

This is also the preparation for looking up into PHI nodes in the
ValueTracker.

rdar://problem/20404526

Differential Revision: http://reviews.llvm.org/D11195

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@242940 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Bruno Cardoso Lopes 2015-07-22 21:30:16 +00:00
parent b18e7bdac8
commit e7aafe5ecc

View File

@ -107,6 +107,8 @@ STATISTIC(NumUncoalescableCopies, "Number of uncoalescable copies optimized");
STATISTIC(NumRewrittenCopies, "Number of copies rewritten"); STATISTIC(NumRewrittenCopies, "Number of copies rewritten");
namespace { namespace {
class ValueTrackerResult;
class PeepholeOptimizer : public MachineFunctionPass { class PeepholeOptimizer : public MachineFunctionPass {
const TargetInstrInfo *TII; const TargetInstrInfo *TII;
const TargetRegisterInfo *TRI; const TargetRegisterInfo *TRI;
@ -171,6 +173,55 @@ namespace {
} }
}; };
/// \brief Helper class to hold a reply for ValueTracker queries. Contains the
/// returned sources for a given search and the instructions where the sources
/// were tracked from.
class ValueTrackerResult {
private:
/// Track all sources found by one ValueTracker query.
SmallVector<TargetInstrInfo::RegSubRegPair, 2> RegSrcs;
/// Instruction using the sources in 'RegSrcs'.
const MachineInstr *Inst;
public:
ValueTrackerResult() : Inst(nullptr) {}
ValueTrackerResult(unsigned Reg, unsigned SubReg) : Inst(nullptr) {
addSource(Reg, SubReg);
}
bool isValid() const { return getNumSources() > 0; }
void setInst(const MachineInstr *I) { Inst = I; }
const MachineInstr *getInst() const { return Inst; }
void clear() {
RegSrcs.clear();
Inst = nullptr;
}
void addSource(unsigned SrcReg, unsigned SrcSubReg) {
RegSrcs.push_back(TargetInstrInfo::RegSubRegPair(SrcReg, SrcSubReg));
}
void setSource(int Idx, unsigned SrcReg, unsigned SrcSubReg) {
assert(Idx < getNumSources() && "Reg pair source out of index");
RegSrcs[Idx] = TargetInstrInfo::RegSubRegPair(SrcReg, SrcSubReg);
}
int getNumSources() const { return RegSrcs.size(); }
unsigned getSrcReg(int Idx) const {
assert(Idx < getNumSources() && "Reg source out of index");
return RegSrcs[Idx].Reg;
}
unsigned getSrcSubReg(int Idx) const {
assert(Idx < getNumSources() && "SubReg source out of index");
return RegSrcs[Idx].SubReg;
}
};
/// \brief Helper class to track the possible sources of a value defined by /// \brief Helper class to track the possible sources of a value defined by
/// a (chain of) copy related instructions. /// a (chain of) copy related instructions.
/// Given a definition (instruction and definition index), this class /// Given a definition (instruction and definition index), this class
@ -213,23 +264,23 @@ namespace {
/// \brief Dispatcher to the right underlying implementation of /// \brief Dispatcher to the right underlying implementation of
/// getNextSource. /// getNextSource.
bool getNextSourceImpl(unsigned &SrcReg, unsigned &SrcSubReg); ValueTrackerResult getNextSourceImpl();
/// \brief Specialized version of getNextSource for Copy instructions. /// \brief Specialized version of getNextSource for Copy instructions.
bool getNextSourceFromCopy(unsigned &SrcReg, unsigned &SrcSubReg); ValueTrackerResult getNextSourceFromCopy();
/// \brief Specialized version of getNextSource for Bitcast instructions. /// \brief Specialized version of getNextSource for Bitcast instructions.
bool getNextSourceFromBitcast(unsigned &SrcReg, unsigned &SrcSubReg); ValueTrackerResult getNextSourceFromBitcast();
/// \brief Specialized version of getNextSource for RegSequence /// \brief Specialized version of getNextSource for RegSequence
/// instructions. /// instructions.
bool getNextSourceFromRegSequence(unsigned &SrcReg, unsigned &SrcSubReg); ValueTrackerResult getNextSourceFromRegSequence();
/// \brief Specialized version of getNextSource for InsertSubreg /// \brief Specialized version of getNextSource for InsertSubreg
/// instructions. /// instructions.
bool getNextSourceFromInsertSubreg(unsigned &SrcReg, unsigned &SrcSubReg); ValueTrackerResult getNextSourceFromInsertSubreg();
/// \brief Specialized version of getNextSource for ExtractSubreg /// \brief Specialized version of getNextSource for ExtractSubreg
/// instructions. /// instructions.
bool getNextSourceFromExtractSubreg(unsigned &SrcReg, unsigned &SrcSubReg); ValueTrackerResult getNextSourceFromExtractSubreg();
/// \brief Specialized version of getNextSource for SubregToReg /// \brief Specialized version of getNextSource for SubregToReg
/// instructions. /// instructions.
bool getNextSourceFromSubregToReg(unsigned &SrcReg, unsigned &SrcSubReg); ValueTrackerResult getNextSourceFromSubregToReg();
public: public:
/// \brief Create a ValueTracker instance for the value defined by \p Reg. /// \brief Create a ValueTracker instance for the value defined by \p Reg.
@ -276,16 +327,10 @@ namespace {
/// \brief Following the use-def chain, get the next available source /// \brief Following the use-def chain, get the next available source
/// for the tracked value. /// for the tracked value.
/// When the returned value is not nullptr, \p SrcReg gives the register /// \return A ValueTrackerResult containing the a set of registers
/// that contain the tracked value. /// and sub registers with tracked values. A ValueTrackerResult with
/// \note The sub register index returned in \p SrcSubReg must be used /// an empty set of registers means no source was found.
/// on \p SrcReg to access the actual value. ValueTrackerResult getNextSource();
/// \return Unless the returned value is nullptr (i.e., no source found),
/// \p SrcReg gives the register of the next source used in the returned
/// instruction and \p SrcSubReg the sub-register index to be used on that
/// source to get the tracked value. When nullptr is returned, no
/// alternative source has been found.
const MachineInstr *getNextSource(unsigned &SrcReg, unsigned &SrcSubReg);
/// \brief Get the last register where the initial value can be found. /// \brief Get the last register where the initial value can be found.
/// Initially this is the register of the definition. /// Initially this is the register of the definition.
@ -560,11 +605,11 @@ bool PeepholeOptimizer::findNextSource(unsigned &Reg, unsigned &SubReg) {
// or find a more suitable source. // or find a more suitable source.
ValueTracker ValTracker(Reg, DefSubReg, *MRI, !DisableAdvCopyOpt, TII); ValueTracker ValTracker(Reg, DefSubReg, *MRI, !DisableAdvCopyOpt, TII);
do { do {
unsigned CopySrcReg, CopySrcSubReg; ValueTrackerResult Res = ValTracker.getNextSource();
if (!ValTracker.getNextSource(CopySrcReg, CopySrcSubReg)) if (!Res.isValid())
break; break;
Src = CopySrcReg; Src = Res.getSrcReg(0);
SrcSubReg = CopySrcSubReg; SrcSubReg = Res.getSrcSubReg(0);
// Do not extend the live-ranges of physical registers as they add // Do not extend the live-ranges of physical registers as they add
// constraints to the register allocator. // constraints to the register allocator.
@ -653,7 +698,7 @@ public:
/// \brief Rewrite the current source with \p NewReg and \p NewSubReg /// \brief Rewrite the current source with \p NewReg and \p NewSubReg
/// if possible. /// if possible.
/// \return True if the rewritting was possible, false otherwise. /// \return True if the rewriting was possible, false otherwise.
virtual bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) { virtual bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) {
if (!CopyLike.isCopy() || CurrentSrcIdx != 1) if (!CopyLike.isCopy() || CurrentSrcIdx != 1)
return false; return false;
@ -662,6 +707,91 @@ public:
MOSrc.setSubReg(NewSubReg); MOSrc.setSubReg(NewSubReg);
return true; return true;
} }
/// \brief Rewrite the current source with \p NewSrcReg and \p NewSecSubReg
/// by creating a new COPY instruction. \p DefReg and \p DefSubReg contain the
/// definition to be rewritten from the original copylike instruction.
/// \return The new COPY if the rewriting was possible, nullptr otherwise.
/// This is needed to handle Uncoalescable copies, since they are copy
/// like instructions that aren't recognized by the register allocator.
virtual MachineInstr *RewriteCurrentSource(unsigned DefReg,
unsigned DefSubReg,
unsigned NewSrcReg,
unsigned NewSrcSubReg) {
return nullptr;
}
};
/// \brief Helper class to rewrite uncoalescable copy like instructions
/// into new COPY (coalescable friendly) instructions.
class UncoalescableRewriter : public CopyRewriter {
protected:
const TargetInstrInfo &TII;
MachineRegisterInfo &MRI;
/// The number of defs in the bitcast
unsigned NumDefs;
public:
UncoalescableRewriter(MachineInstr &MI, const TargetInstrInfo &TII,
MachineRegisterInfo &MRI)
: CopyRewriter(MI), TII(TII), MRI(MRI) {
NumDefs = MI.getDesc().getNumDefs();
}
/// \brief Get the next rewritable def source (TrackReg, TrackSubReg)
/// All such sources need to be considered rewritable in order to
/// rewrite a uncoalescable copy-like instruction. This method return
/// each definition that must be checked if rewritable.
///
bool getNextRewritableSource(unsigned &SrcReg, unsigned &SrcSubReg,
unsigned &TrackReg,
unsigned &TrackSubReg) override {
// Find the next non-dead definition and continue from there.
if (CurrentSrcIdx == NumDefs)
return false;
while (CopyLike.getOperand(CurrentSrcIdx).isDead()) {
++CurrentSrcIdx;
if (CurrentSrcIdx == NumDefs)
return false;
}
// What we track are the alternative sources of the definition.
const MachineOperand &MODef = CopyLike.getOperand(CurrentSrcIdx);
TrackReg = MODef.getReg();
TrackSubReg = MODef.getSubReg();
CurrentSrcIdx++;
return true;
}
/// \brief Rewrite the current source with \p NewSrcReg and \p NewSrcSubReg
/// by creating a new COPY instruction. \p DefReg and \p DefSubReg contain the
/// definition to be rewritten from the original copylike instruction.
/// \return The new COPY if the rewriting was possible, nullptr otherwise.
MachineInstr *RewriteCurrentSource(unsigned DefReg, unsigned DefSubReg,
unsigned NewSrcReg,
unsigned NewSrcSubReg) override {
assert(!TargetRegisterInfo::isPhysicalRegister(DefReg) &&
"We do not rewrite physical registers");
const TargetRegisterClass *DefRC = MRI.getRegClass(DefReg);
unsigned NewVR = MRI.createVirtualRegister(DefRC);
MachineInstr *NewCopy =
BuildMI(*CopyLike.getParent(), &CopyLike, CopyLike.getDebugLoc(),
TII.get(TargetOpcode::COPY), NewVR)
.addReg(NewSrcReg, 0, NewSrcSubReg);
NewCopy->getOperand(0).setSubReg(DefSubReg);
if (DefSubReg)
NewCopy->getOperand(0).setIsUndef();
MRI.replaceRegWith(DefReg, NewVR);
MRI.clearKillFlags(NewVR);
return NewCopy;
}
}; };
/// \brief Specialized rewriter for INSERT_SUBREG instruction. /// \brief Specialized rewriter for INSERT_SUBREG instruction.
@ -850,7 +980,13 @@ public:
/// \return A pointer to a dynamically allocated CopyRewriter or nullptr /// \return A pointer to a dynamically allocated CopyRewriter or nullptr
/// if no rewriter works for \p MI. /// if no rewriter works for \p MI.
static CopyRewriter *getCopyRewriter(MachineInstr &MI, static CopyRewriter *getCopyRewriter(MachineInstr &MI,
const TargetInstrInfo &TII) { const TargetInstrInfo &TII,
MachineRegisterInfo &MRI) {
// Handle uncoalescable copy-like instructions.
if (MI.isBitcast() || (MI.isRegSequenceLike() || MI.isInsertSubregLike() ||
MI.isExtractSubregLike()))
return new UncoalescableRewriter(MI, TII, MRI);
switch (MI.getOpcode()) { switch (MI.getOpcode()) {
default: default:
return nullptr; return nullptr;
@ -889,7 +1025,7 @@ bool PeepholeOptimizer::optimizeCoalescableCopy(MachineInstr *MI) {
bool Changed = false; bool Changed = false;
// Get the right rewriter for the current copy. // Get the right rewriter for the current copy.
std::unique_ptr<CopyRewriter> CpyRewriter(getCopyRewriter(*MI, *TII)); std::unique_ptr<CopyRewriter> CpyRewriter(getCopyRewriter(*MI, *TII, *MRI));
// If none exists, bails out. // If none exists, bails out.
if (!CpyRewriter) if (!CpyRewriter)
return false; return false;
@ -899,9 +1035,8 @@ bool PeepholeOptimizer::optimizeCoalescableCopy(MachineInstr *MI) {
TrackSubReg)) { TrackSubReg)) {
unsigned NewSrc = TrackReg; unsigned NewSrc = TrackReg;
unsigned NewSubReg = TrackSubReg; unsigned NewSubReg = TrackSubReg;
// Try to find a more suitable source. // Try to find a more suitable source. If we failed to do so, or get the
// If we failed to do so, or get the actual source, // actual source, move to the next source.
// move to the next source.
if (!findNextSource(NewSrc, NewSubReg) || SrcReg == NewSrc) if (!findNextSource(NewSrc, NewSubReg) || SrcReg == NewSrc)
continue; continue;
// Rewrite source. // Rewrite source.
@ -939,45 +1074,47 @@ bool PeepholeOptimizer::optimizeUncoalescableCopy(
SmallVector< SmallVector<
std::pair<TargetInstrInfo::RegSubRegPair, TargetInstrInfo::RegSubRegPair>, std::pair<TargetInstrInfo::RegSubRegPair, TargetInstrInfo::RegSubRegPair>,
4> RewritePairs; 4> RewritePairs;
for (const MachineOperand &MODef : MI->defs()) { // Get the right rewriter for the current copy.
if (MODef.isDead()) std::unique_ptr<CopyRewriter> CpyRewriter(getCopyRewriter(*MI, *TII, *MRI));
// We can ignore those. // If none exists, bails out.
continue; if (!CpyRewriter)
return false;
// Rewrite each rewritable source by generating new COPYs. This works
// differently from optimizeCoalescableCopy since it first makes sure that all
// definitions can be rewritten.
unsigned SrcReg, SrcSubReg, TrackReg, TrackSubReg;
while (CpyRewriter->getNextRewritableSource(SrcReg, SrcSubReg, TrackReg,
TrackSubReg)) {
// If a physical register is here, this is probably for a good reason. // If a physical register is here, this is probably for a good reason.
// Do not rewrite that. // Do not rewrite that.
if (TargetRegisterInfo::isPhysicalRegister(MODef.getReg())) if (TargetRegisterInfo::isPhysicalRegister(TrackReg))
return false; return false;
// If we do not know how to rewrite this definition, there is no point // If we do not know how to rewrite this definition, there is no point
// in trying to kill this instruction. // in trying to kill this instruction.
TargetInstrInfo::RegSubRegPair Def(MODef.getReg(), MODef.getSubReg()); TargetInstrInfo::RegSubRegPair Def(TrackReg, TrackSubReg);
TargetInstrInfo::RegSubRegPair Src = Def; TargetInstrInfo::RegSubRegPair Src = Def;
if (!findNextSource(Src.Reg, Src.SubReg)) if (!findNextSource(Src.Reg, Src.SubReg))
return false; return false;
RewritePairs.push_back(std::make_pair(Def, Src)); RewritePairs.push_back(std::make_pair(Def, Src));
} }
// The change is possible for all defs, do it. // The change is possible for all defs, do it.
for (const auto &PairDefSrc : RewritePairs) { for (const auto &PairDefSrc : RewritePairs) {
const auto &Def = PairDefSrc.first; const auto &Def = PairDefSrc.first;
const auto &Src = PairDefSrc.second; const auto &Src = PairDefSrc.second;
// Rewrite the "copy" in a way the register coalescer understands. // Rewrite the "copy" in a way the register coalescer understands.
assert(!TargetRegisterInfo::isPhysicalRegister(Def.Reg) && MachineInstr *NewCopy = CpyRewriter->RewriteCurrentSource(
"We do not rewrite physical registers"); Def.Reg, Def.SubReg, Src.Reg, Src.SubReg);
const TargetRegisterClass *DefRC = MRI->getRegClass(Def.Reg); assert(NewCopy && "Should be able to always generate a new copy");
unsigned NewVR = MRI->createVirtualRegister(DefRC);
MachineInstr *NewCopy = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), // We extended the lifetime of Src and clear the kill flags to
TII->get(TargetOpcode::COPY), // account for that.
NewVR).addReg(Src.Reg, 0, Src.SubReg);
NewCopy->getOperand(0).setSubReg(Def.SubReg);
if (Def.SubReg)
NewCopy->getOperand(0).setIsUndef();
LocalMIs.insert(NewCopy);
MRI->replaceRegWith(Def.Reg, NewVR);
MRI->clearKillFlags(NewVR);
// We extended the lifetime of Src.
// Clear the kill flags to account for that.
MRI->clearKillFlags(Src.Reg); MRI->clearKillFlags(Src.Reg);
LocalMIs.insert(NewCopy);
} }
// MI is now dead. // MI is now dead.
MI->eraseFromParent(); MI->eraseFromParent();
@ -1190,8 +1327,7 @@ bool PeepholeOptimizer::runOnMachineFunction(MachineFunction &MF) {
return Changed; return Changed;
} }
bool ValueTracker::getNextSourceFromCopy(unsigned &SrcReg, ValueTrackerResult ValueTracker::getNextSourceFromCopy() {
unsigned &SrcSubReg) {
assert(Def->isCopy() && "Invalid definition"); assert(Def->isCopy() && "Invalid definition");
// Copy instruction are supposed to be: Def = Src. // Copy instruction are supposed to be: Def = Src.
// If someone breaks this assumption, bad things will happen everywhere. // If someone breaks this assumption, bad things will happen everywhere.
@ -1200,29 +1336,26 @@ bool ValueTracker::getNextSourceFromCopy(unsigned &SrcReg,
if (Def->getOperand(DefIdx).getSubReg() != DefSubReg) if (Def->getOperand(DefIdx).getSubReg() != DefSubReg)
// If we look for a different subreg, it means we want a subreg of src. // If we look for a different subreg, it means we want a subreg of src.
// Bails as we do not support composing subreg yet. // Bails as we do not support composing subreg yet.
return false; return ValueTrackerResult();
// Otherwise, we want the whole source. // Otherwise, we want the whole source.
const MachineOperand &Src = Def->getOperand(1); const MachineOperand &Src = Def->getOperand(1);
SrcReg = Src.getReg(); return ValueTrackerResult(Src.getReg(), Src.getSubReg());
SrcSubReg = Src.getSubReg();
return true;
} }
bool ValueTracker::getNextSourceFromBitcast(unsigned &SrcReg, ValueTrackerResult ValueTracker::getNextSourceFromBitcast() {
unsigned &SrcSubReg) {
assert(Def->isBitcast() && "Invalid definition"); assert(Def->isBitcast() && "Invalid definition");
// Bail if there are effects that a plain copy will not expose. // Bail if there are effects that a plain copy will not expose.
if (Def->hasUnmodeledSideEffects()) if (Def->hasUnmodeledSideEffects())
return false; return ValueTrackerResult();
// Bitcasts with more than one def are not supported. // Bitcasts with more than one def are not supported.
if (Def->getDesc().getNumDefs() != 1) if (Def->getDesc().getNumDefs() != 1)
return false; return ValueTrackerResult();
if (Def->getOperand(DefIdx).getSubReg() != DefSubReg) if (Def->getOperand(DefIdx).getSubReg() != DefSubReg)
// If we look for a different subreg, it means we want a subreg of the src. // If we look for a different subreg, it means we want a subreg of the src.
// Bails as we do not support composing subreg yet. // Bails as we do not support composing subreg yet.
return false; return ValueTrackerResult();
unsigned SrcIdx = Def->getNumOperands(); unsigned SrcIdx = Def->getNumOperands();
for (unsigned OpIdx = DefIdx + 1, EndOpIdx = SrcIdx; OpIdx != EndOpIdx; for (unsigned OpIdx = DefIdx + 1, EndOpIdx = SrcIdx; OpIdx != EndOpIdx;
@ -1233,17 +1366,14 @@ bool ValueTracker::getNextSourceFromBitcast(unsigned &SrcReg,
assert(!MO.isDef() && "We should have skipped all the definitions by now"); assert(!MO.isDef() && "We should have skipped all the definitions by now");
if (SrcIdx != EndOpIdx) if (SrcIdx != EndOpIdx)
// Multiple sources? // Multiple sources?
return false; return ValueTrackerResult();
SrcIdx = OpIdx; SrcIdx = OpIdx;
} }
const MachineOperand &Src = Def->getOperand(SrcIdx); const MachineOperand &Src = Def->getOperand(SrcIdx);
SrcReg = Src.getReg(); return ValueTrackerResult(Src.getReg(), Src.getSubReg());
SrcSubReg = Src.getSubReg();
return true;
} }
bool ValueTracker::getNextSourceFromRegSequence(unsigned &SrcReg, ValueTrackerResult ValueTracker::getNextSourceFromRegSequence() {
unsigned &SrcSubReg) {
assert((Def->isRegSequence() || Def->isRegSequenceLike()) && assert((Def->isRegSequence() || Def->isRegSequenceLike()) &&
"Invalid definition"); "Invalid definition");
@ -1262,16 +1392,16 @@ bool ValueTracker::getNextSourceFromRegSequence(unsigned &SrcReg,
// have this case. // have this case.
// If we can ascertain (or force) that this never happens, we could // If we can ascertain (or force) that this never happens, we could
// turn that into an assertion. // turn that into an assertion.
return false; return ValueTrackerResult();
if (!TII) if (!TII)
// We could handle the REG_SEQUENCE here, but we do not want to // We could handle the REG_SEQUENCE here, but we do not want to
// duplicate the code from the generic TII. // duplicate the code from the generic TII.
return false; return ValueTrackerResult();
SmallVector<TargetInstrInfo::RegSubRegPairAndIdx, 8> RegSeqInputRegs; SmallVector<TargetInstrInfo::RegSubRegPairAndIdx, 8> RegSeqInputRegs;
if (!TII->getRegSequenceInputs(*Def, DefIdx, RegSeqInputRegs)) if (!TII->getRegSequenceInputs(*Def, DefIdx, RegSeqInputRegs))
return false; return ValueTrackerResult();
// We are looking at: // We are looking at:
// Def = REG_SEQUENCE v0, sub0, v1, sub1, ... // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
@ -1280,22 +1410,19 @@ bool ValueTracker::getNextSourceFromRegSequence(unsigned &SrcReg,
if (RegSeqInput.SubIdx == DefSubReg) { if (RegSeqInput.SubIdx == DefSubReg) {
if (RegSeqInput.SubReg) if (RegSeqInput.SubReg)
// Bails if we have to compose sub registers. // Bails if we have to compose sub registers.
return false; return ValueTrackerResult();
SrcReg = RegSeqInput.Reg; return ValueTrackerResult(RegSeqInput.Reg, RegSeqInput.SubReg);
SrcSubReg = RegSeqInput.SubReg;
return true;
} }
} }
// If the subreg we are tracking is super-defined by another subreg, // If the subreg we are tracking is super-defined by another subreg,
// we could follow this value. However, this would require to compose // we could follow this value. However, this would require to compose
// the subreg and we do not do that for now. // the subreg and we do not do that for now.
return false; return ValueTrackerResult();
} }
bool ValueTracker::getNextSourceFromInsertSubreg(unsigned &SrcReg, ValueTrackerResult ValueTracker::getNextSourceFromInsertSubreg() {
unsigned &SrcSubReg) {
assert((Def->isInsertSubreg() || Def->isInsertSubregLike()) && assert((Def->isInsertSubreg() || Def->isInsertSubregLike()) &&
"Invalid definition"); "Invalid definition");
@ -1303,17 +1430,17 @@ bool ValueTracker::getNextSourceFromInsertSubreg(unsigned &SrcReg,
// If we are composing subreg, bails out. // If we are composing subreg, bails out.
// Same remark as getNextSourceFromRegSequence. // Same remark as getNextSourceFromRegSequence.
// I.e., this may be turned into an assert. // I.e., this may be turned into an assert.
return false; return ValueTrackerResult();
if (!TII) if (!TII)
// We could handle the REG_SEQUENCE here, but we do not want to // We could handle the REG_SEQUENCE here, but we do not want to
// duplicate the code from the generic TII. // duplicate the code from the generic TII.
return false; return ValueTrackerResult();
TargetInstrInfo::RegSubRegPair BaseReg; TargetInstrInfo::RegSubRegPair BaseReg;
TargetInstrInfo::RegSubRegPairAndIdx InsertedReg; TargetInstrInfo::RegSubRegPairAndIdx InsertedReg;
if (!TII->getInsertSubregInputs(*Def, DefIdx, BaseReg, InsertedReg)) if (!TII->getInsertSubregInputs(*Def, DefIdx, BaseReg, InsertedReg))
return false; return ValueTrackerResult();
// We are looking at: // We are looking at:
// Def = INSERT_SUBREG v0, v1, sub1 // Def = INSERT_SUBREG v0, v1, sub1
@ -1323,9 +1450,7 @@ bool ValueTracker::getNextSourceFromInsertSubreg(unsigned &SrcReg,
// #1 Check if the inserted register matches the required sub index. // #1 Check if the inserted register matches the required sub index.
if (InsertedReg.SubIdx == DefSubReg) { if (InsertedReg.SubIdx == DefSubReg) {
SrcReg = InsertedReg.Reg; return ValueTrackerResult(InsertedReg.Reg, InsertedReg.SubReg);
SrcSubReg = InsertedReg.SubReg;
return true;
} }
// #2 Otherwise, if the sub register we are looking for is not partial // #2 Otherwise, if the sub register we are looking for is not partial
// defined by the inserted element, we can look through the main // defined by the inserted element, we can look through the main
@ -1336,7 +1461,7 @@ bool ValueTracker::getNextSourceFromInsertSubreg(unsigned &SrcReg,
// subregisters, bails out. // subregisters, bails out.
if (MRI.getRegClass(MODef.getReg()) != MRI.getRegClass(BaseReg.Reg) || if (MRI.getRegClass(MODef.getReg()) != MRI.getRegClass(BaseReg.Reg) ||
BaseReg.SubReg) BaseReg.SubReg)
return false; return ValueTrackerResult();
// Get the TRI and check if the inserted sub-register overlaps with the // Get the TRI and check if the inserted sub-register overlaps with the
// sub-register we are tracking. // sub-register we are tracking.
@ -1344,16 +1469,13 @@ bool ValueTracker::getNextSourceFromInsertSubreg(unsigned &SrcReg,
if (!TRI || if (!TRI ||
(TRI->getSubRegIndexLaneMask(DefSubReg) & (TRI->getSubRegIndexLaneMask(DefSubReg) &
TRI->getSubRegIndexLaneMask(InsertedReg.SubIdx)) != 0) TRI->getSubRegIndexLaneMask(InsertedReg.SubIdx)) != 0)
return false; return ValueTrackerResult();
// At this point, the value is available in v0 via the same subreg // At this point, the value is available in v0 via the same subreg
// we used for Def. // we used for Def.
SrcReg = BaseReg.Reg; return ValueTrackerResult(BaseReg.Reg, DefSubReg);
SrcSubReg = DefSubReg;
return true;
} }
bool ValueTracker::getNextSourceFromExtractSubreg(unsigned &SrcReg, ValueTrackerResult ValueTracker::getNextSourceFromExtractSubreg() {
unsigned &SrcSubReg) {
assert((Def->isExtractSubreg() || assert((Def->isExtractSubreg() ||
Def->isExtractSubregLike()) && "Invalid definition"); Def->isExtractSubregLike()) && "Invalid definition");
// We are looking at: // We are looking at:
@ -1362,29 +1484,26 @@ bool ValueTracker::getNextSourceFromExtractSubreg(unsigned &SrcReg,
// Bails if we have to compose sub registers. // Bails if we have to compose sub registers.
// Indeed, if DefSubReg != 0, we would have to compose it with sub0. // Indeed, if DefSubReg != 0, we would have to compose it with sub0.
if (DefSubReg) if (DefSubReg)
return false; return ValueTrackerResult();
if (!TII) if (!TII)
// We could handle the EXTRACT_SUBREG here, but we do not want to // We could handle the EXTRACT_SUBREG here, but we do not want to
// duplicate the code from the generic TII. // duplicate the code from the generic TII.
return false; return ValueTrackerResult();
TargetInstrInfo::RegSubRegPairAndIdx ExtractSubregInputReg; TargetInstrInfo::RegSubRegPairAndIdx ExtractSubregInputReg;
if (!TII->getExtractSubregInputs(*Def, DefIdx, ExtractSubregInputReg)) if (!TII->getExtractSubregInputs(*Def, DefIdx, ExtractSubregInputReg))
return false; return ValueTrackerResult();
// Bails if we have to compose sub registers. // Bails if we have to compose sub registers.
// Likewise, if v0.subreg != 0, we would have to compose v0.subreg with sub0. // Likewise, if v0.subreg != 0, we would have to compose v0.subreg with sub0.
if (ExtractSubregInputReg.SubReg) if (ExtractSubregInputReg.SubReg)
return false; return ValueTrackerResult();
// Otherwise, the value is available in the v0.sub0. // Otherwise, the value is available in the v0.sub0.
SrcReg = ExtractSubregInputReg.Reg; return ValueTrackerResult(ExtractSubregInputReg.Reg, ExtractSubregInputReg.SubIdx);
SrcSubReg = ExtractSubregInputReg.SubIdx;
return true;
} }
bool ValueTracker::getNextSourceFromSubregToReg(unsigned &SrcReg, ValueTrackerResult ValueTracker::getNextSourceFromSubregToReg() {
unsigned &SrcSubReg) {
assert(Def->isSubregToReg() && "Invalid definition"); assert(Def->isSubregToReg() && "Invalid definition");
// We are looking at: // We are looking at:
// Def = SUBREG_TO_REG Imm, v0, sub0 // Def = SUBREG_TO_REG Imm, v0, sub0
@ -1394,71 +1513,71 @@ bool ValueTracker::getNextSourceFromSubregToReg(unsigned &SrcReg,
// we track are included in sub0 and if yes, we would have to // we track are included in sub0 and if yes, we would have to
// determine the right subreg in v0. // determine the right subreg in v0.
if (DefSubReg != Def->getOperand(3).getImm()) if (DefSubReg != Def->getOperand(3).getImm())
return false; return ValueTrackerResult();
// Bails if we have to compose sub registers. // Bails if we have to compose sub registers.
// Likewise, if v0.subreg != 0, we would have to compose it with sub0. // Likewise, if v0.subreg != 0, we would have to compose it with sub0.
if (Def->getOperand(2).getSubReg()) if (Def->getOperand(2).getSubReg())
return false; return ValueTrackerResult();
SrcReg = Def->getOperand(2).getReg(); return ValueTrackerResult(Def->getOperand(2).getReg(),
SrcSubReg = Def->getOperand(3).getImm(); Def->getOperand(3).getImm());
return true;
} }
bool ValueTracker::getNextSourceImpl(unsigned &SrcReg, unsigned &SrcSubReg) { ValueTrackerResult ValueTracker::getNextSourceImpl() {
assert(Def && "This method needs a valid definition"); assert(Def && "This method needs a valid definition");
assert( assert(
(DefIdx < Def->getDesc().getNumDefs() || Def->getDesc().isVariadic()) && (DefIdx < Def->getDesc().getNumDefs() || Def->getDesc().isVariadic()) &&
Def->getOperand(DefIdx).isDef() && "Invalid DefIdx"); Def->getOperand(DefIdx).isDef() && "Invalid DefIdx");
if (Def->isCopy()) if (Def->isCopy())
return getNextSourceFromCopy(SrcReg, SrcSubReg); return getNextSourceFromCopy();
if (Def->isBitcast()) if (Def->isBitcast())
return getNextSourceFromBitcast(SrcReg, SrcSubReg); return getNextSourceFromBitcast();
// All the remaining cases involve "complex" instructions. // All the remaining cases involve "complex" instructions.
// Bails if we did not ask for the advanced tracking. // Bails if we did not ask for the advanced tracking.
if (!UseAdvancedTracking) if (!UseAdvancedTracking)
return false; return ValueTrackerResult();
if (Def->isRegSequence() || Def->isRegSequenceLike()) if (Def->isRegSequence() || Def->isRegSequenceLike())
return getNextSourceFromRegSequence(SrcReg, SrcSubReg); return getNextSourceFromRegSequence();
if (Def->isInsertSubreg() || Def->isInsertSubregLike()) if (Def->isInsertSubreg() || Def->isInsertSubregLike())
return getNextSourceFromInsertSubreg(SrcReg, SrcSubReg); return getNextSourceFromInsertSubreg();
if (Def->isExtractSubreg() || Def->isExtractSubregLike()) if (Def->isExtractSubreg() || Def->isExtractSubregLike())
return getNextSourceFromExtractSubreg(SrcReg, SrcSubReg); return getNextSourceFromExtractSubreg();
if (Def->isSubregToReg()) if (Def->isSubregToReg())
return getNextSourceFromSubregToReg(SrcReg, SrcSubReg); return getNextSourceFromSubregToReg();
return false; return ValueTrackerResult();
} }
const MachineInstr *ValueTracker::getNextSource(unsigned &SrcReg, ValueTrackerResult ValueTracker::getNextSource() {
unsigned &SrcSubReg) {
// If we reach a point where we cannot move up in the use-def chain, // If we reach a point where we cannot move up in the use-def chain,
// there is nothing we can get. // there is nothing we can get.
if (!Def) if (!Def)
return nullptr; return ValueTrackerResult();
const MachineInstr *PrevDef = nullptr; ValueTrackerResult Res = getNextSourceImpl();
// Try to find the next source. if (Res.isValid()) {
if (getNextSourceImpl(SrcReg, SrcSubReg)) {
// Update definition, definition index, and subregister for the // Update definition, definition index, and subregister for the
// next call of getNextSource. // next call of getNextSource.
// Update the current register. // Update the current register.
Reg = SrcReg; bool OneRegSrc = Res.getNumSources() == 1;
// Update the return value before moving up in the use-def chain. if (OneRegSrc)
PrevDef = Def; Reg = Res.getSrcReg(0);
// Update the result before moving up in the use-def chain
// with the instruction containing the last found sources.
Res.setInst(Def);
// If we can still move up in the use-def chain, move to the next // If we can still move up in the use-def chain, move to the next
// defintion. // defintion.
if (!TargetRegisterInfo::isPhysicalRegister(Reg)) { if (!TargetRegisterInfo::isPhysicalRegister(Reg) && OneRegSrc) {
Def = MRI.getVRegDef(Reg); Def = MRI.getVRegDef(Reg);
DefIdx = MRI.def_begin(Reg).getOperandNo(); DefIdx = MRI.def_begin(Reg).getOperandNo();
DefSubReg = SrcSubReg; DefSubReg = Res.getSrcSubReg(0);
return PrevDef; return Res;
} }
} }
// If we end up here, this means we will not be able to find another source // If we end up here, this means we will not be able to find another source
// for the next iteration. // for the next iteration. Make sure any new call to getNextSource bails out
// Make sure any new call to getNextSource bails out early by cutting the // early by cutting the use-def chain.
// use-def chain.
Def = nullptr; Def = nullptr;
return PrevDef; return Res;
} }