/// SelectionDAG operations.
///
class VISIBILITY_HIDDEN X86DAGToDAGISel : public SelectionDAGISel {
- /// TM - Keep a reference to X86TargetMachine.
- ///
- X86TargetMachine &TM;
-
/// X86Lowering - This object fully describes how to lower LLVM code to an
/// X86-specific SelectionDAG.
X86TargetLowering &X86Lowering;
bool OptForSize;
public:
- explicit X86DAGToDAGISel(X86TargetMachine &tm, unsigned OptLevel)
+ explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
: SelectionDAGISel(tm, OptLevel),
- TM(tm), X86Lowering(*TM.getTargetLowering()),
- Subtarget(&TM.getSubtarget<X86Subtarget>()),
+ X86Lowering(*tm.getTargetLowering()),
+ Subtarget(&tm.getSubtarget<X86Subtarget>()),
OptForSize(false) {}
virtual const char *getPassName() const {
SDValue &Segment);
bool SelectLEAAddr(SDValue Op, SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index, SDValue &Disp);
+ bool SelectTLSADDRAddr(SDValue Op, SDValue N, SDValue &Base,
+ SDValue &Scale, SDValue &Index, SDValue &Disp);
bool SelectScalarSSELoad(SDValue Op, SDValue Pred,
SDValue N, SDValue &Base, SDValue &Scale,
SDValue &Index, SDValue &Disp,
///
SDNode *getGlobalBaseReg();
+ /// getTargetMachine - Return a reference to the TargetMachine, casted
+ /// to the target-specific type.
+ const X86TargetMachine &getTargetMachine() {
+ return static_cast<const X86TargetMachine &>(TM);
+ }
+
+ /// getInstrInfo - Return a reference to the TargetInstrInfo, casted
+ /// to the target-specific type.
+ const X86InstrInfo *getInstrInfo() {
+ return getTargetMachine().getInstrInfo();
+ }
+
#ifndef NDEBUG
unsigned Indent;
#endif
};
}
-/// findFlagUse - Return use of MVT::Flag value produced by the specified
-/// SDNode.
-///
-static SDNode *findFlagUse(SDNode *N) {
- unsigned FlagResNo = N->getNumValues()-1;
- for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
- SDUse &Use = I.getUse();
- if (Use.getResNo() == FlagResNo)
- return Use.getUser();
- }
- return NULL;
-}
-
-/// findNonImmUse - Return true if "Use" is a non-immediate use of "Def".
-/// This function recursively traverses up the operand chain, ignoring
-/// certain nodes.
-static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
- SDNode *Root,
- SmallPtrSet<SDNode*, 16> &Visited) {
- if (Use->getNodeId() < Def->getNodeId() ||
- !Visited.insert(Use))
- return false;
-
- for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
- SDNode *N = Use->getOperand(i).getNode();
- if (N == Def) {
- if (Use == ImmedUse || Use == Root)
- continue; // We are not looking for immediate use.
- assert(N != Root);
- return true;
- }
-
- // Traverse up the operand chain.
- if (findNonImmUse(N, Def, ImmedUse, Root, Visited))
- return true;
- }
- return false;
-}
-
-/// isNonImmUse - Start searching from Root up the DAG to check is Def can
-/// be reached. Return true if that's the case. However, ignore direct uses
-/// by ImmedUse (which would be U in the example illustrated in
-/// IsLegalAndProfitableToFold) and by Root (which can happen in the store
-/// case).
-/// FIXME: to be really generic, we should allow direct use by any node
-/// that is being folded. But realisticly since we only fold loads which
-/// have one non-chain use, we only need to watch out for load/op/store
-/// and load/op/cmp case where the root (store / cmp) may reach the load via
-/// its chain operand.
-static inline bool isNonImmUse(SDNode *Root, SDNode *Def, SDNode *ImmedUse) {
- SmallPtrSet<SDNode*, 16> Visited;
- return findNonImmUse(Root, Def, ImmedUse, Root, Visited);
-}
-
bool X86DAGToDAGISel::IsLegalAndProfitableToFold(SDNode *N, SDNode *U,
SDNode *Root) const {
- if (OptLevel == 0) return false;
+ if (OptLevel == CodeGenOpt::None) return false;
if (U == Root)
switch (U->getOpcode()) {
}
}
- // If Root use can somehow reach N through a path that that doesn't contain
- // U then folding N would create a cycle. e.g. In the following
- // diagram, Root can reach N through X. If N is folded into into Root, then
- // X is both a predecessor and a successor of U.
- //
- // [N*] //
- // ^ ^ //
- // / \ //
- // [U*] [X]? //
- // ^ ^ //
- // \ / //
- // \ / //
- // [Root*] //
- //
- // * indicates nodes to be folded together.
- //
- // If Root produces a flag, then it gets (even more) interesting. Since it
- // will be "glued" together with its flag use in the scheduler, we need to
- // check if it might reach N.
- //
- // [N*] //
- // ^ ^ //
- // / \ //
- // [U*] [X]? //
- // ^ ^ //
- // \ \ //
- // \ | //
- // [Root*] | //
- // ^ | //
- // f | //
- // | / //
- // [Y] / //
- // ^ / //
- // f / //
- // | / //
- // [FU] //
- //
- // If FU (flag use) indirectly reaches N (the load), and Root folds N
- // (call it Fold), then X is a predecessor of FU and a successor of
- // Fold. But since Fold and FU are flagged together, this will create
- // a cycle in the scheduling graph.
-
- MVT VT = Root->getValueType(Root->getNumValues()-1);
- while (VT == MVT::Flag) {
- SDNode *FU = findFlagUse(Root);
- if (FU == NULL)
- break;
- Root = FU;
- VT = Root->getValueType(Root->getNumValues()-1);
- }
-
- return !isNonImmUse(Root, N, U);
+ // Proceed to 'generic' cycle finder code
+ return SelectionDAGISel::IsLegalAndProfitableToFold(N, U, Root);
}
/// MoveBelowTokenFactor - Replace TokenFactor operand with load's chain operand
OptForSize = F->hasFnAttr(Attribute::OptimizeForSize);
DEBUG(BB->dump());
- if (OptLevel != 0)
+ if (OptLevel != CodeGenOpt::None)
PreprocessForRMW();
// FIXME: This should only happen when not compiled with -O0.
}
bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
+ bool SymbolicAddressesAreRIPRel =
+ getTargetMachine().symbolicAddressesAreRIPRel();
bool is64Bit = Subtarget->is64Bit();
DOUT << "Wrapper: 64bit " << is64Bit;
DOUT << " AM "; DEBUG(AM.dump()); DOUT << "\n";
// Base and index reg must be 0 in order to use rip as base.
bool canUsePICRel = !AM.Base.Reg.getNode() && !AM.IndexReg.getNode();
- if (is64Bit && !canUsePICRel && TM.symbolicAddressesAreRIPRel())
+ if (is64Bit && !canUsePICRel && SymbolicAddressesAreRIPRel)
return true;
if (AM.hasSymbolicDisplacement())
uint64_t Offset = G->getOffset();
if (!is64Bit || isInt32(AM.Disp + Offset)) {
GlobalValue *GV = G->getGlobal();
- bool isRIPRel = TM.symbolicAddressesAreRIPRel();
+ bool isRIPRel = SymbolicAddressesAreRIPRel;
if (N0.getOpcode() == llvm::ISD::TargetGlobalTLSAddress) {
TLSModel::Model model =
getTLSModel (GV, TM.getRelocationModel());
AM.CP = CP->getConstVal();
AM.Align = CP->getAlignment();
AM.Disp += Offset;
- AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
+ AM.isRIPRel = SymbolicAddressesAreRIPRel;
return false;
}
} else if (ExternalSymbolSDNode *S =dyn_cast<ExternalSymbolSDNode>(N0)) {
AM.ES = S->getSymbol();
- AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
+ AM.isRIPRel = SymbolicAddressesAreRIPRel;
return false;
} else if (JumpTableSDNode *J = dyn_cast<JumpTableSDNode>(N0)) {
AM.JT = J->getIndex();
- AM.isRIPRel = TM.symbolicAddressesAreRIPRel();
+ AM.isRIPRel = SymbolicAddressesAreRIPRel;
return false;
}
}
break;
+ case ISD::SUB: {
+ // Given A-B, if A can be completely folded into the address and
+ // the index field with the index field unused, use -B as the index.
+ // This is a win if a has multiple parts that can be folded into
+ // the address. Also, this saves a mov if the base register has
+ // other uses, since it avoids a two-address sub instruction, however
+ // it costs an additional mov if the index register has other uses.
+
+ // Test if the LHS of the sub can be folded.
+ X86ISelAddressMode Backup = AM;
+ if (MatchAddress(N.getNode()->getOperand(0), AM, Depth+1)) {
+ AM = Backup;
+ break;
+ }
+ // Test if the index field is free for use.
+ if (AM.IndexReg.getNode() || AM.isRIPRel) {
+ AM = Backup;
+ break;
+ }
+ int Cost = 0;
+ SDValue RHS = N.getNode()->getOperand(1);
+ // If the RHS involves a register with multiple uses, this
+ // transformation incurs an extra mov, due to the neg instruction
+ // clobbering its operand.
+ if (!RHS.getNode()->hasOneUse() ||
+ RHS.getNode()->getOpcode() == ISD::CopyFromReg ||
+ RHS.getNode()->getOpcode() == ISD::TRUNCATE ||
+ RHS.getNode()->getOpcode() == ISD::ANY_EXTEND ||
+ (RHS.getNode()->getOpcode() == ISD::ZERO_EXTEND &&
+ RHS.getNode()->getOperand(0).getValueType() == MVT::i32))
+ ++Cost;
+ // If the base is a register with multiple uses, this
+ // transformation may save a mov.
+ if ((AM.BaseType == X86ISelAddressMode::RegBase &&
+ AM.Base.Reg.getNode() &&
+ !AM.Base.Reg.getNode()->hasOneUse()) ||
+ AM.BaseType == X86ISelAddressMode::FrameIndexBase)
+ --Cost;
+ // If the folded LHS was interesting, this transformation saves
+ // address arithmetic.
+ if ((AM.hasSymbolicDisplacement() && !Backup.hasSymbolicDisplacement()) +
+ ((AM.Disp != 0) && (Backup.Disp == 0)) +
+ (AM.Segment.getNode() && !Backup.Segment.getNode()) >= 2)
+ --Cost;
+ // If it doesn't look like it may be an overall win, don't do it.
+ if (Cost >= 0) {
+ AM = Backup;
+ break;
+ }
+
+ // Ok, the transformation is legal and appears profitable. Go for it.
+ SDValue Zero = CurDAG->getConstant(0, N.getValueType());
+ SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
+ AM.IndexReg = Neg;
+ AM.Scale = 1;
+
+ // Insert the new nodes into the topological ordering.
+ if (Zero.getNode()->getNodeId() == -1 ||
+ Zero.getNode()->getNodeId() > N.getNode()->getNodeId()) {
+ CurDAG->RepositionNode(N.getNode(), Zero.getNode());
+ Zero.getNode()->setNodeId(N.getNode()->getNodeId());
+ }
+ if (Neg.getNode()->getNodeId() == -1 ||
+ Neg.getNode()->getNodeId() > N.getNode()->getNodeId()) {
+ CurDAG->RepositionNode(N.getNode(), Neg.getNode());
+ Neg.getNode()->setNodeId(N.getNode()->getNodeId());
+ }
+ return false;
+ }
+
case ISD::ADD: {
X86ISelAddressMode Backup = AM;
if (!MatchAddress(N.getNode()->getOperand(0), AM, Depth+1) &&
return false;
}
+/// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
+bool X86DAGToDAGISel::SelectTLSADDRAddr(SDValue Op, SDValue N, SDValue &Base,
+ SDValue &Scale, SDValue &Index,
+ SDValue &Disp) {
+ assert(Op.getOpcode() == X86ISD::TLSADDR);
+ assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
+ const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
+
+ X86ISelAddressMode AM;
+ AM.GV = GA->getGlobal();
+ AM.Disp += GA->getOffset();
+ AM.Base.Reg = CurDAG->getRegister(0, N.getValueType());
+
+ if (N.getValueType() == MVT::i32) {
+ AM.Scale = 1;
+ AM.IndexReg = CurDAG->getRegister(X86::EBX, MVT::i32);
+ } else {
+ AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
+ }
+
+ SDValue Segment;
+ getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
+ return true;
+}
+
+
bool X86DAGToDAGISel::TryFoldLoad(SDValue P, SDValue N,
SDValue &Base, SDValue &Scale,
SDValue &Index, SDValue &Disp,
///
SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
MachineFunction *MF = CurBB->getParent();
- unsigned GlobalBaseReg = TM.getInstrInfo()->getGlobalBaseReg(MF);
+ unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
}
/// createX86ISelDag - This pass converts a legalized DAG into a
/// X86-specific DAG, ready for instruction scheduling.
///
-FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM, unsigned OptLevel) {
+FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
+ llvm::CodeGenOpt::Level OptLevel) {
return new X86DAGToDAGISel(TM, OptLevel);
}
projects / oota-llvm.git / blobdiff