//===----------------------------------------------------------------------===//
namespace {
- /// X86ISelAddressMode - This corresponds to X86AddressMode, but uses
- /// SDValue's instead of register numbers for the leaves of the matched
- /// tree.
+ /// This corresponds to X86AddressMode, but uses SDValue's instead of register
+ /// numbers for the leaves of the matched tree.
struct X86ISelAddressMode {
enum {
RegBase,
IndexReg.getNode() != nullptr || Base_Reg.getNode() != nullptr;
}
- /// isRIPRelative - Return true if this addressing mode is already RIP
- /// relative.
+ /// Return true if this addressing mode is already RIP-relative.
bool isRIPRelative() const {
if (BaseType != RegBase) return false;
if (RegisterSDNode *RegNode =
namespace {
//===--------------------------------------------------------------------===//
- /// ISel - X86 specific code to select X86 machine instructions for
+ /// ISel - X86-specific code to select X86 machine instructions for
/// SelectionDAG operations.
///
class X86DAGToDAGISel final : public SelectionDAGISel {
- /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
+ /// Keep a pointer to the X86Subtarget around so that we can
/// make the right decision when generating code for different targets.
const X86Subtarget *Subtarget;
- /// OptForSize - If true, selector should try to optimize for code size
- /// instead of performance.
+ /// If true, selector should try to optimize for code size instead of
+ /// performance.
bool OptForSize;
public:
return isInt<8>(cast<ConstantSDNode>(N)->getSExtValue());
}
- // i64immSExt32 predicate - True if the 64-bit immediate fits in a 32-bit
- // sign extended field.
+ // True if the 64-bit immediate fits in a 32-bit sign-extended field.
inline bool i64immSExt32(SDNode *N) const {
uint64_t v = cast<ConstantSDNode>(N)->getZExtValue();
return (int64_t)v == (int32_t)v;
private:
SDNode *Select(SDNode *N) override;
- SDNode *SelectGather(SDNode *N, unsigned Opc);
- SDNode *SelectAtomicLoadArith(SDNode *Node, MVT NVT);
-
- bool FoldOffsetIntoAddress(uint64_t Offset, X86ISelAddressMode &AM);
- bool MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM);
- bool MatchWrapper(SDValue N, X86ISelAddressMode &AM);
- bool MatchAddress(SDValue N, X86ISelAddressMode &AM);
- bool MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
+ SDNode *selectGather(SDNode *N, unsigned Opc);
+ SDNode *selectAtomicLoadArith(SDNode *Node, MVT NVT);
+
+ bool foldOffsetIntoAddress(uint64_t Offset, X86ISelAddressMode &AM);
+ bool matchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM);
+ bool matchWrapper(SDValue N, X86ISelAddressMode &AM);
+ bool matchAddress(SDValue N, X86ISelAddressMode &AM);
+ bool matchAdd(SDValue N, X86ISelAddressMode &AM, unsigned Depth);
+ bool matchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
unsigned Depth);
- bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
- bool SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
+ bool matchAddressBase(SDValue N, X86ISelAddressMode &AM);
+ bool selectAddr(SDNode *Parent, SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index, SDValue &Disp,
SDValue &Segment);
- bool SelectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
+ bool selectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index, SDValue &Disp,
SDValue &Segment);
- bool SelectMOV64Imm32(SDValue N, SDValue &Imm);
- bool SelectLEAAddr(SDValue N, SDValue &Base,
+ bool selectMOV64Imm32(SDValue N, SDValue &Imm);
+ bool selectLEAAddr(SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index, SDValue &Disp,
SDValue &Segment);
- bool SelectLEA64_32Addr(SDValue N, SDValue &Base,
+ bool selectLEA64_32Addr(SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index, SDValue &Disp,
SDValue &Segment);
- bool SelectTLSADDRAddr(SDValue N, SDValue &Base,
+ bool selectTLSADDRAddr(SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index, SDValue &Disp,
SDValue &Segment);
- bool SelectScalarSSELoad(SDNode *Root, SDValue N,
+ bool selectScalarSSELoad(SDNode *Root, SDValue N,
SDValue &Base, SDValue &Scale,
SDValue &Index, SDValue &Disp,
SDValue &Segment,
SDValue &NodeWithChain);
- bool TryFoldLoad(SDNode *P, SDValue N,
+ bool tryFoldLoad(SDNode *P, SDValue N,
SDValue &Base, SDValue &Scale,
SDValue &Index, SDValue &Disp,
SDValue &Segment);
- /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
- /// inline asm expressions.
+ /// Implement addressing mode selection for inline asm expressions.
bool SelectInlineAsmMemoryOperand(const SDValue &Op,
unsigned ConstraintID,
std::vector<SDValue> &OutOps) override;
- void EmitSpecialCodeForMain();
+ void emitSpecialCodeForMain();
inline void getAddressOperands(X86ISelAddressMode &AM, SDLoc DL,
SDValue &Base, SDValue &Scale,
: AM.Base_Reg;
Scale = getI8Imm(AM.Scale, DL);
Index = AM.IndexReg;
- // These are 32-bit even in 64-bit mode since RIP relative offset
+ // These are 32-bit even in 64-bit mode since RIP-relative offset
// is 32-bit.
if (AM.GV)
Disp = CurDAG->getTargetGlobalAddress(AM.GV, SDLoc(),
return (UseCount > 1);
}
- /// getI8Imm - Return a target constant with the specified value, of type
- /// i8.
+ /// Return a target constant with the specified value of type i8.
inline SDValue getI8Imm(unsigned Imm, SDLoc DL) {
return CurDAG->getTargetConstant(Imm, DL, MVT::i8);
}
- /// getI32Imm - Return a target constant with the specified value, of type
- /// i32.
+ /// Return a target constant with the specified value, of type i32.
inline SDValue getI32Imm(unsigned Imm, SDLoc DL) {
return CurDAG->getTargetConstant(Imm, DL, MVT::i32);
}
- /// getGlobalBaseReg - Return an SDNode that returns the value of
- /// the global base register. Output instructions required to
- /// initialize the global base register, if necessary.
- ///
+ /// Return an SDNode that returns the value of the global base register.
+ /// Output instructions required to initialize the global base register,
+ /// if necessary.
SDNode *getGlobalBaseReg();
- /// getTargetMachine - Return a reference to the TargetMachine, casted
- /// to the target-specific type.
+ /// Return a reference to the TargetMachine, casted to the target-specific
+ /// type.
const X86TargetMachine &getTargetMachine() const {
return static_cast<const X86TargetMachine &>(TM);
}
- /// getInstrInfo - Return a reference to the TargetInstrInfo, casted
- /// to the target-specific type.
+ /// Return a reference to the TargetInstrInfo, casted to the target-specific
+ /// type.
const X86InstrInfo *getInstrInfo() const {
return Subtarget->getInstrInfo();
}
return true;
}
-/// MoveBelowCallOrigChain - Replace the original chain operand of the call with
+/// Replace the original chain operand of the call with
/// load's chain operand and move load below the call's chain operand.
-static void MoveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
+static void moveBelowOrigChain(SelectionDAG *CurDAG, SDValue Load,
SDValue Call, SDValue OrigChain) {
SmallVector<SDValue, 8> Ops;
SDValue Chain = OrigChain.getOperand(0);
CurDAG->UpdateNodeOperands(Call.getNode(), Ops);
}
-/// isCalleeLoad - Return true if call address is a load and it can be
+/// Return true if call address is a load and it can be
/// moved below CALLSEQ_START and the chains leading up to the call.
/// Return the CALLSEQ_START by reference as a second output.
/// In the case of a tail call, there isn't a callseq node between the call
for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
E = CurDAG->allnodes_end(); I != E; ) {
- SDNode *N = I++; // Preincrement iterator to avoid invalidation issues.
+ SDNode *N = &*I++; // Preincrement iterator to avoid invalidation issues.
if (OptLevel != CodeGenOpt::None &&
// Only does this when target favors doesn't favor register indirect
SDValue Load = N->getOperand(1);
if (!isCalleeLoad(Load, Chain, HasCallSeq))
continue;
- MoveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
+ moveBelowOrigChain(CurDAG, Load, SDValue(N, 0), Chain);
++NumLoadMoved;
continue;
}
}
-/// EmitSpecialCodeForMain - Emit any code that needs to be executed only in
-/// the main function.
-void X86DAGToDAGISel::EmitSpecialCodeForMain() {
+/// Emit any code that needs to be executed only in the main function.
+void X86DAGToDAGISel::emitSpecialCodeForMain() {
if (Subtarget->isTargetCygMing()) {
TargetLowering::ArgListTy Args;
auto &DL = CurDAG->getDataLayout();
// If this is main, emit special code for main.
if (const Function *Fn = MF->getFunction())
if (Fn->hasExternalLinkage() && Fn->getName() == "main")
- EmitSpecialCodeForMain();
+ emitSpecialCodeForMain();
}
static bool isDispSafeForFrameIndex(int64_t Val) {
return isInt<31>(Val);
}
-bool X86DAGToDAGISel::FoldOffsetIntoAddress(uint64_t Offset,
+bool X86DAGToDAGISel::foldOffsetIntoAddress(uint64_t Offset,
X86ISelAddressMode &AM) {
// Cannot combine ExternalSymbol displacements with integer offsets.
if (Offset != 0 && (AM.ES || AM.MCSym))
}
-bool X86DAGToDAGISel::MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM){
+bool X86DAGToDAGISel::matchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM){
SDValue Address = N->getOperand(1);
// load gs:0 -> GS segment register.
return true;
}
-/// MatchWrapper - Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes
-/// into an addressing mode. These wrap things that will resolve down into a
-/// symbol reference. If no match is possible, this returns true, otherwise it
-/// returns false.
-bool X86DAGToDAGISel::MatchWrapper(SDValue N, X86ISelAddressMode &AM) {
+/// Try to match X86ISD::Wrapper and X86ISD::WrapperRIP nodes into an addressing
+/// mode. These wrap things that will resolve down into a symbol reference.
+/// If no match is possible, this returns true, otherwise it returns false.
+bool X86DAGToDAGISel::matchWrapper(SDValue N, X86ISelAddressMode &AM) {
// If the addressing mode already has a symbol as the displacement, we can
// never match another symbol.
if (AM.hasSymbolicDisplacement())
X86ISelAddressMode Backup = AM;
AM.GV = G->getGlobal();
AM.SymbolFlags = G->getTargetFlags();
- if (FoldOffsetIntoAddress(G->getOffset(), AM)) {
+ if (foldOffsetIntoAddress(G->getOffset(), AM)) {
AM = Backup;
return true;
}
AM.CP = CP->getConstVal();
AM.Align = CP->getAlignment();
AM.SymbolFlags = CP->getTargetFlags();
- if (FoldOffsetIntoAddress(CP->getOffset(), AM)) {
+ if (foldOffsetIntoAddress(CP->getOffset(), AM)) {
AM = Backup;
return true;
}
X86ISelAddressMode Backup = AM;
AM.BlockAddr = BA->getBlockAddress();
AM.SymbolFlags = BA->getTargetFlags();
- if (FoldOffsetIntoAddress(BA->getOffset(), AM)) {
+ if (foldOffsetIntoAddress(BA->getOffset(), AM)) {
AM = Backup;
return true;
}
return true;
}
-/// MatchAddress - Add the specified node to the specified addressing mode,
-/// returning true if it cannot be done. This just pattern matches for the
-/// addressing mode.
-bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
- if (MatchAddressRecursively(N, AM, 0))
+/// Add the specified node to the specified addressing mode, returning true if
+/// it cannot be done. This just pattern matches for the addressing mode.
+bool X86DAGToDAGISel::matchAddress(SDValue N, X86ISelAddressMode &AM) {
+ if (matchAddressRecursively(N, AM, 0))
return true;
// Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
return false;
}
+bool X86DAGToDAGISel::matchAdd(SDValue N, X86ISelAddressMode &AM,
+ unsigned Depth) {
+ // Add an artificial use to this node so that we can keep track of
+ // it if it gets CSE'd with a different node.
+ HandleSDNode Handle(N);
+
+ X86ISelAddressMode Backup = AM;
+ if (!matchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
+ !matchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1))
+ return false;
+ AM = Backup;
+
+ // Try again after commuting the operands.
+ if (!matchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1) &&
+ !matchAddressRecursively(Handle.getValue().getOperand(0), AM, Depth+1))
+ return false;
+ AM = Backup;
+
+ // If we couldn't fold both operands into the address at the same time,
+ // see if we can just put each operand into a register and fold at least
+ // the add.
+ if (AM.BaseType == X86ISelAddressMode::RegBase &&
+ !AM.Base_Reg.getNode() &&
+ !AM.IndexReg.getNode()) {
+ N = Handle.getValue();
+ AM.Base_Reg = N.getOperand(0);
+ AM.IndexReg = N.getOperand(1);
+ AM.Scale = 1;
+ return false;
+ }
+ N = Handle.getValue();
+ return true;
+}
+
// Insert a node into the DAG at least before the Pos node's position. This
// will reposition the node as needed, and will assign it a node ID that is <=
// the Pos node's ID. Note that this does *not* preserve the uniqueness of node
// IDs! The selection DAG must no longer depend on their uniqueness when this
// is used.
-static void InsertDAGNode(SelectionDAG &DAG, SDValue Pos, SDValue N) {
+static void insertDAGNode(SelectionDAG &DAG, SDValue Pos, SDValue N) {
if (N.getNode()->getNodeId() == -1 ||
N.getNode()->getNodeId() > Pos.getNode()->getNodeId()) {
- DAG.RepositionNode(Pos.getNode(), N.getNode());
+ DAG.RepositionNode(Pos.getNode()->getIterator(), N.getNode());
N.getNode()->setNodeId(Pos.getNode()->getNodeId());
}
}
// safe. This allows us to convert the shift and and into an h-register
// extract and a scaled index. Returns false if the simplification is
// performed.
-static bool FoldMaskAndShiftToExtract(SelectionDAG &DAG, SDValue N,
+static bool foldMaskAndShiftToExtract(SelectionDAG &DAG, SDValue N,
uint64_t Mask,
SDValue Shift, SDValue X,
X86ISelAddressMode &AM) {
// these nodes. We continually insert before 'N' in sequence as this is
// essentially a pre-flattened and pre-sorted sequence of nodes. There is no
// hierarchy left to express.
- InsertDAGNode(DAG, N, Eight);
- InsertDAGNode(DAG, N, Srl);
- InsertDAGNode(DAG, N, NewMask);
- InsertDAGNode(DAG, N, And);
- InsertDAGNode(DAG, N, ShlCount);
- InsertDAGNode(DAG, N, Shl);
+ insertDAGNode(DAG, N, Eight);
+ insertDAGNode(DAG, N, Srl);
+ insertDAGNode(DAG, N, NewMask);
+ insertDAGNode(DAG, N, And);
+ insertDAGNode(DAG, N, ShlCount);
+ insertDAGNode(DAG, N, Shl);
DAG.ReplaceAllUsesWith(N, Shl);
AM.IndexReg = And;
AM.Scale = (1 << ScaleLog);
// Transforms "(X << C1) & C2" to "(X & (C2>>C1)) << C1" if safe and if this
// allows us to fold the shift into this addressing mode. Returns false if the
// transform succeeded.
-static bool FoldMaskedShiftToScaledMask(SelectionDAG &DAG, SDValue N,
+static bool foldMaskedShiftToScaledMask(SelectionDAG &DAG, SDValue N,
uint64_t Mask,
SDValue Shift, SDValue X,
X86ISelAddressMode &AM) {
// these nodes. We continually insert before 'N' in sequence as this is
// essentially a pre-flattened and pre-sorted sequence of nodes. There is no
// hierarchy left to express.
- InsertDAGNode(DAG, N, NewMask);
- InsertDAGNode(DAG, N, NewAnd);
- InsertDAGNode(DAG, N, NewShift);
+ insertDAGNode(DAG, N, NewMask);
+ insertDAGNode(DAG, N, NewAnd);
+ insertDAGNode(DAG, N, NewShift);
DAG.ReplaceAllUsesWith(N, NewShift);
AM.Scale = 1 << ShiftAmt;
// Note that this function assumes the mask is provided as a mask *after* the
// value is shifted. The input chain may or may not match that, but computing
// such a mask is trivial.
-static bool FoldMaskAndShiftToScale(SelectionDAG &DAG, SDValue N,
+static bool foldMaskAndShiftToScale(SelectionDAG &DAG, SDValue N,
uint64_t Mask,
SDValue Shift, SDValue X,
X86ISelAddressMode &AM) {
assert(X.getValueType() != VT);
// We looked through an ANY_EXTEND node, insert a ZERO_EXTEND.
SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(X), VT, X);
- InsertDAGNode(DAG, N, NewX);
+ insertDAGNode(DAG, N, NewX);
X = NewX;
}
SDLoc DL(N);
// these nodes. We continually insert before 'N' in sequence as this is
// essentially a pre-flattened and pre-sorted sequence of nodes. There is no
// hierarchy left to express.
- InsertDAGNode(DAG, N, NewSRLAmt);
- InsertDAGNode(DAG, N, NewSRL);
- InsertDAGNode(DAG, N, NewSHLAmt);
- InsertDAGNode(DAG, N, NewSHL);
+ insertDAGNode(DAG, N, NewSRLAmt);
+ insertDAGNode(DAG, N, NewSRL);
+ insertDAGNode(DAG, N, NewSHLAmt);
+ insertDAGNode(DAG, N, NewSHL);
DAG.ReplaceAllUsesWith(N, NewSHL);
AM.Scale = 1 << AMShiftAmt;
return false;
}
-bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
+bool X86DAGToDAGISel::matchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
unsigned Depth) {
SDLoc dl(N);
DEBUG({
});
// Limit recursion.
if (Depth > 5)
- return MatchAddressBase(N, AM);
+ return matchAddressBase(N, AM);
// If this is already a %rip relative address, we can only merge immediates
// into it. Instead of handling this in every case, we handle it here.
return true;
if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N))
- if (!FoldOffsetIntoAddress(Cst->getSExtValue(), AM))
+ if (!foldOffsetIntoAddress(Cst->getSExtValue(), AM))
return false;
return true;
}
}
case ISD::Constant: {
uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
- if (!FoldOffsetIntoAddress(Val, AM))
+ if (!foldOffsetIntoAddress(Val, AM))
return false;
break;
}
case X86ISD::Wrapper:
case X86ISD::WrapperRIP:
- if (!MatchWrapper(N, AM))
+ if (!matchWrapper(N, AM))
return false;
break;
case ISD::LOAD:
- if (!MatchLoadInAddress(cast<LoadSDNode>(N), AM))
+ if (!matchLoadInAddress(cast<LoadSDNode>(N), AM))
return false;
break;
ConstantSDNode *AddVal =
cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
uint64_t Disp = (uint64_t)AddVal->getSExtValue() << Val;
- if (!FoldOffsetIntoAddress(Disp, AM))
+ if (!foldOffsetIntoAddress(Disp, AM))
return false;
}
// Try to fold the mask and shift into the scale, and return false if we
// succeed.
- if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, N, X, AM))
+ if (!foldMaskAndShiftToScale(*CurDAG, N, Mask, N, X, AM))
return false;
break;
}
ConstantSDNode *AddVal =
cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
uint64_t Disp = AddVal->getSExtValue() * CN->getZExtValue();
- if (FoldOffsetIntoAddress(Disp, AM))
+ if (foldOffsetIntoAddress(Disp, AM))
Reg = N.getNode()->getOperand(0);
} else {
Reg = N.getNode()->getOperand(0);
// Test if the LHS of the sub can be folded.
X86ISelAddressMode Backup = AM;
- if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
+ if (matchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
AM = Backup;
break;
}
AM.Scale = 1;
// Insert the new nodes into the topological ordering.
- InsertDAGNode(*CurDAG, N, Zero);
- InsertDAGNode(*CurDAG, N, Neg);
+ insertDAGNode(*CurDAG, N, Zero);
+ insertDAGNode(*CurDAG, N, Neg);
return false;
}
- case ISD::ADD: {
- // Add an artificial use to this node so that we can keep track of
- // it if it gets CSE'd with a different node.
- HandleSDNode Handle(N);
-
- X86ISelAddressMode Backup = AM;
- if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
- !MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1))
- return false;
- AM = Backup;
-
- // Try again after commuting the operands.
- if (!MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1)&&
- !MatchAddressRecursively(Handle.getValue().getOperand(0), AM, Depth+1))
+ case ISD::ADD:
+ if (!matchAdd(N, AM, Depth))
return false;
- AM = Backup;
-
- // If we couldn't fold both operands into the address at the same time,
- // see if we can just put each operand into a register and fold at least
- // the add.
- if (AM.BaseType == X86ISelAddressMode::RegBase &&
- !AM.Base_Reg.getNode() &&
- !AM.IndexReg.getNode()) {
- N = Handle.getValue();
- AM.Base_Reg = N.getOperand(0);
- AM.IndexReg = N.getOperand(1);
- AM.Scale = 1;
- return false;
- }
- N = Handle.getValue();
break;
- }
case ISD::OR:
- // Handle "X | C" as "X + C" iff X is known to have C bits clear.
- if (CurDAG->isBaseWithConstantOffset(N)) {
- X86ISelAddressMode Backup = AM;
- ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
-
- // Start with the LHS as an addr mode.
- if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
- !FoldOffsetIntoAddress(CN->getSExtValue(), AM))
- return false;
- AM = Backup;
- }
+ // We want to look through a transform in InstCombine and DAGCombiner that
+ // turns 'add' into 'or', so we can treat this 'or' exactly like an 'add'.
+ // Example: (or (and x, 1), (shl y, 3)) --> (add (and x, 1), (shl y, 3))
+ // An 'lea' can then be used to match the shift (multiply) and add:
+ // and $1, %esi
+ // lea (%rsi, %rdi, 8), %rax
+ if (CurDAG->haveNoCommonBitsSet(N.getOperand(0), N.getOperand(1)) &&
+ !matchAdd(N, AM, Depth))
+ return false;
break;
case ISD::AND: {
uint64_t Mask = N.getConstantOperandVal(1);
// Try to fold the mask and shift into an extract and scale.
- if (!FoldMaskAndShiftToExtract(*CurDAG, N, Mask, Shift, X, AM))
+ if (!foldMaskAndShiftToExtract(*CurDAG, N, Mask, Shift, X, AM))
return false;
// Try to fold the mask and shift directly into the scale.
- if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, Shift, X, AM))
+ if (!foldMaskAndShiftToScale(*CurDAG, N, Mask, Shift, X, AM))
return false;
// Try to swap the mask and shift to place shifts which can be done as
// a scale on the outside of the mask.
- if (!FoldMaskedShiftToScaledMask(*CurDAG, N, Mask, Shift, X, AM))
+ if (!foldMaskedShiftToScaledMask(*CurDAG, N, Mask, Shift, X, AM))
return false;
break;
}
}
- return MatchAddressBase(N, AM);
+ return matchAddressBase(N, AM);
}
-/// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
+/// Helper for MatchAddress. Add the specified node to the
/// specified addressing mode without any further recursion.
-bool X86DAGToDAGISel::MatchAddressBase(SDValue N, X86ISelAddressMode &AM) {
+bool X86DAGToDAGISel::matchAddressBase(SDValue N, X86ISelAddressMode &AM) {
// Is the base register already occupied?
if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base_Reg.getNode()) {
// If so, check to see if the scale index register is set.
return false;
}
-bool X86DAGToDAGISel::SelectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
+bool X86DAGToDAGISel::selectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index,
SDValue &Disp, SDValue &Segment) {
// If Base is 0, the whole address is in index and the Scale is 1
if (isa<ConstantSDNode>(Base)) {
- assert(dyn_cast<ConstantSDNode>(Base)->isNullValue() &&
+ assert(cast<ConstantSDNode>(Base)->isNullValue() &&
"Unexpected base in gather/scatter");
Scale = getI8Imm(1, DL);
Base = CurDAG->getRegister(0, MVT::i32);
return true;
}
-/// SelectAddr - returns true if it is able pattern match an addressing mode.
+/// Returns true if it is able to pattern match an addressing mode.
/// It returns the operands which make up the maximal addressing mode it can
/// match by reference.
///
/// Parent is the parent node of the addr operand that is being matched. It
/// is always a load, store, atomic node, or null. It is only null when
/// checking memory operands for inline asm nodes.
-bool X86DAGToDAGISel::SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
+bool X86DAGToDAGISel::selectAddr(SDNode *Parent, SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index,
SDValue &Disp, SDValue &Segment) {
X86ISelAddressMode AM;
AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
}
- if (MatchAddress(N, AM))
+ if (matchAddress(N, AM))
return false;
MVT VT = N.getSimpleValueType();
return true;
}
-/// SelectScalarSSELoad - Match a scalar SSE load. In particular, we want to
-/// match a load whose top elements are either undef or zeros. The load flavor
-/// is derived from the type of N, which is either v4f32 or v2f64.
+/// Match a scalar SSE load. In particular, we want to match a load whose top
+/// elements are either undef or zeros. The load flavor is derived from the
+/// type of N, which is either v4f32 or v2f64.
///
/// We also return:
/// PatternChainNode: this is the matched node that has a chain input and
/// output.
-bool X86DAGToDAGISel::SelectScalarSSELoad(SDNode *Root,
+bool X86DAGToDAGISel::selectScalarSSELoad(SDNode *Root,
SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index,
SDValue &Disp, SDValue &Segment,
IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
- if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
+ if (!selectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
return false;
return true;
}
IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
// Okay, this is a zero extending load. Fold it.
LoadSDNode *LD = cast<LoadSDNode>(N.getOperand(0).getOperand(0));
- if (!SelectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
+ if (!selectAddr(LD, LD->getBasePtr(), Base, Scale, Index, Disp, Segment))
return false;
PatternNodeWithChain = SDValue(LD, 0);
return true;
}
-bool X86DAGToDAGISel::SelectMOV64Imm32(SDValue N, SDValue &Imm) {
+bool X86DAGToDAGISel::selectMOV64Imm32(SDValue N, SDValue &Imm) {
if (const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
uint64_t ImmVal = CN->getZExtValue();
if ((uint32_t)ImmVal != (uint64_t)ImmVal)
return TM.getCodeModel() == CodeModel::Small;
}
-bool X86DAGToDAGISel::SelectLEA64_32Addr(SDValue N, SDValue &Base,
+bool X86DAGToDAGISel::selectLEA64_32Addr(SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index,
SDValue &Disp, SDValue &Segment) {
- if (!SelectLEAAddr(N, Base, Scale, Index, Disp, Segment))
+ if (!selectLEAAddr(N, Base, Scale, Index, Disp, Segment))
return false;
SDLoc DL(N);
return true;
}
-/// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
+/// Calls SelectAddr and determines if the maximal addressing
/// mode it matches can be cost effectively emitted as an LEA instruction.
-bool X86DAGToDAGISel::SelectLEAAddr(SDValue N,
+bool X86DAGToDAGISel::selectLEAAddr(SDValue N,
SDValue &Base, SDValue &Scale,
SDValue &Index, SDValue &Disp,
SDValue &Segment) {
SDValue Copy = AM.Segment;
SDValue T = CurDAG->getRegister(0, MVT::i32);
AM.Segment = T;
- if (MatchAddress(N, AM))
+ if (matchAddress(N, AM))
return false;
assert (T == AM.Segment);
AM.Segment = Copy;
Complexity++;
// FIXME: We are artificially lowering the criteria to turn ADD %reg, $GA
- // to a LEA. This is determined with some expermentation but is by no means
+ // to a LEA. This is determined with some experimentation but is by no means
// optimal (especially for code size consideration). LEA is nice because of
// its three-address nature. Tweak the cost function again when we can run
// convertToThreeAddress() at register allocation time.
if (AM.hasSymbolicDisplacement()) {
- // For X86-64, we should always use lea to materialize RIP relative
- // addresses.
+ // For X86-64, always use LEA to materialize RIP-relative addresses.
if (Subtarget->is64Bit())
Complexity = 4;
else
return true;
}
-/// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
-bool X86DAGToDAGISel::SelectTLSADDRAddr(SDValue N, SDValue &Base,
+/// This is only run on TargetGlobalTLSAddress nodes.
+bool X86DAGToDAGISel::selectTLSADDRAddr(SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index,
SDValue &Disp, SDValue &Segment) {
assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
}
-bool X86DAGToDAGISel::TryFoldLoad(SDNode *P, SDValue N,
+bool X86DAGToDAGISel::tryFoldLoad(SDNode *P, SDValue N,
SDValue &Base, SDValue &Scale,
SDValue &Index, SDValue &Disp,
SDValue &Segment) {
!IsLegalToFold(N, P, P, OptLevel))
return false;
- return SelectAddr(N.getNode(),
+ return selectAddr(N.getNode(),
N.getOperand(1), Base, Scale, Index, Disp, Segment);
}
-/// getGlobalBaseReg - Return an SDNode that returns the value of
-/// the global base register. Output instructions required to
-/// initialize the global base register, if necessary.
-///
+/// Return an SDNode that returns the value of the global base register.
+/// Output instructions required to initialize the global base register,
+/// if necessary.
SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
auto &DL = MF->getDataLayout();
return Val;
}
-SDNode *X86DAGToDAGISel::SelectAtomicLoadArith(SDNode *Node, MVT NVT) {
+SDNode *X86DAGToDAGISel::selectAtomicLoadArith(SDNode *Node, MVT NVT) {
if (Node->hasAnyUseOfValue(0))
return nullptr;
SDValue Ptr = Node->getOperand(1);
SDValue Val = Node->getOperand(2);
SDValue Base, Scale, Index, Disp, Segment;
- if (!SelectAddr(Node, Ptr, Base, Scale, Index, Disp, Segment))
+ if (!selectAddr(Node, Ptr, Base, Scale, Index, Disp, Segment))
return nullptr;
// Which index into the table.
return CurDAG->getMergeValues(RetVals, dl).getNode();
}
-/// HasNoSignedComparisonUses - Test whether the given X86ISD::CMP node has
-/// any uses which require the SF or OF bits to be accurate.
-static bool HasNoSignedComparisonUses(SDNode *N) {
+/// Test whether the given X86ISD::CMP node has any uses which require the SF
+/// or OF bits to be accurate.
+static bool hasNoSignedComparisonUses(SDNode *N) {
// Examine each user of the node.
for (SDNode::use_iterator UI = N->use_begin(),
UE = N->use_end(); UI != UE; ++UI) {
return true;
}
-/// isLoadIncOrDecStore - Check whether or not the chain ending in StoreNode
-/// is suitable for doing the {load; increment or decrement; store} to modify
-/// transformation.
+/// Check whether or not the chain ending in StoreNode is suitable for doing
+/// the {load; increment or decrement; store} to modify transformation.
static bool isLoadIncOrDecStore(StoreSDNode *StoreNode, unsigned Opc,
SDValue StoredVal, SelectionDAG *CurDAG,
LoadSDNode* &LoadNode, SDValue &InputChain) {
return true;
}
-/// getFusedLdStOpcode - Get the appropriate X86 opcode for an in memory
-/// increment or decrement. Opc should be X86ISD::DEC or X86ISD::INC.
+/// Get the appropriate X86 opcode for an in-memory increment or decrement.
+/// Opc should be X86ISD::DEC or X86ISD::INC.
static unsigned getFusedLdStOpcode(EVT &LdVT, unsigned Opc) {
if (Opc == X86ISD::DEC) {
if (LdVT == MVT::i64) return X86::DEC64m;
llvm_unreachable("unrecognized size for LdVT");
}
-/// SelectGather - Customized ISel for GATHER operations.
-///
-SDNode *X86DAGToDAGISel::SelectGather(SDNode *Node, unsigned Opc) {
+/// Customized ISel for GATHER operations.
+SDNode *X86DAGToDAGISel::selectGather(SDNode *Node, unsigned Opc) {
// Operands of Gather: VSrc, Base, VIdx, VMask, Scale
SDValue Chain = Node->getOperand(0);
SDValue VSrc = Node->getOperand(2);
switch (Opcode) {
default: break;
+ case ISD::BRIND: {
+ if (Subtarget->isTargetNaCl())
+ // NaCl has its own pass where jmp %r32 are converted to jmp %r64. We
+ // leave the instruction alone.
+ break;
+ if (Subtarget->isTarget64BitILP32()) {
+ // Converts a 32-bit register to a 64-bit, zero-extended version of
+ // it. This is needed because x86-64 can do many things, but jmp %r32
+ // ain't one of them.
+ const SDValue &Target = Node->getOperand(1);
+ assert(Target.getSimpleValueType() == llvm::MVT::i32);
+ SDValue ZextTarget = CurDAG->getZExtOrTrunc(Target, dl, EVT(MVT::i64));
+ SDValue Brind = CurDAG->getNode(ISD::BRIND, dl, MVT::Other,
+ Node->getOperand(0), ZextTarget);
+ ReplaceUses(SDValue(Node, 0), Brind);
+ SelectCode(ZextTarget.getNode());
+ SelectCode(Brind.getNode());
+ return nullptr;
+ }
+ break;
+ }
case ISD::INTRINSIC_W_CHAIN: {
unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
switch (IntNo) {
case Intrinsic::x86_avx2_gather_q_d: Opc = X86::VPGATHERQDrm; break;
case Intrinsic::x86_avx2_gather_q_d_256: Opc = X86::VPGATHERQDYrm; break;
}
- SDNode *RetVal = SelectGather(Node, Opc);
+ SDNode *RetVal = selectGather(Node, Opc);
if (RetVal)
// We already called ReplaceUses inside SelectGather.
return nullptr;
case ISD::ATOMIC_LOAD_AND:
case ISD::ATOMIC_LOAD_OR:
case ISD::ATOMIC_LOAD_ADD: {
- SDNode *RetVal = SelectAtomicLoadArith(Node, NVT);
+ SDNode *RetVal = selectAtomicLoadArith(Node, NVT);
if (RetVal)
return RetVal;
break;
}
SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
- bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
+ bool foldedLoad = tryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
// Multiply is commmutative.
if (!foldedLoad) {
- foldedLoad = TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
+ foldedLoad = tryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
if (foldedLoad)
std::swap(N0, N1);
}
}
SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
- bool foldedLoad = TryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
+ bool foldedLoad = tryFoldLoad(Node, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
bool signBitIsZero = CurDAG->SignBitIsZero(N0);
SDValue InFlag;
// Special case for div8, just use a move with zero extension to AX to
// clear the upper 8 bits (AH).
SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
- if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
+ if (tryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
Move =
SDValue(CurDAG->getMachineNode(X86::MOVZX32rm8, dl, MVT::i32,
SDValue N1 = Node->getOperand(1);
if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
- HasNoSignedComparisonUses(Node))
+ hasNoSignedComparisonUses(Node))
N0 = N0.getOperand(0);
// Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
// For example, convert "testl %eax, $8" to "testb %al, $8"
if ((C->getZExtValue() & ~UINT64_C(0xff)) == 0 &&
(!(C->getZExtValue() & 0x80) ||
- HasNoSignedComparisonUses(Node))) {
+ hasNoSignedComparisonUses(Node))) {
SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl, MVT::i8);
SDValue Reg = N0.getNode()->getOperand(0);
// For example, "testl %eax, $2048" to "testb %ah, $8".
if ((C->getZExtValue() & ~UINT64_C(0xff00)) == 0 &&
(!(C->getZExtValue() & 0x8000) ||
- HasNoSignedComparisonUses(Node))) {
+ hasNoSignedComparisonUses(Node))) {
// Shift the immediate right by 8 bits.
SDValue ShiftedImm = CurDAG->getTargetConstant(C->getZExtValue() >> 8,
dl, MVT::i8);
if ((C->getZExtValue() & ~UINT64_C(0xffff)) == 0 &&
N0.getValueType() != MVT::i16 &&
(!(C->getZExtValue() & 0x8000) ||
- HasNoSignedComparisonUses(Node))) {
+ hasNoSignedComparisonUses(Node))) {
SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl,
MVT::i16);
SDValue Reg = N0.getNode()->getOperand(0);
if ((C->getZExtValue() & ~UINT64_C(0xffffffff)) == 0 &&
N0.getValueType() == MVT::i64 &&
(!(C->getZExtValue() & 0x80000000) ||
- HasNoSignedComparisonUses(Node))) {
+ hasNoSignedComparisonUses(Node))) {
SDValue Imm = CurDAG->getTargetConstant(C->getZExtValue(), dl,
MVT::i32);
SDValue Reg = N0.getNode()->getOperand(0);
break;
SDValue Base, Scale, Index, Disp, Segment;
- if (!SelectAddr(LoadNode, LoadNode->getBasePtr(),
+ if (!selectAddr(LoadNode, LoadNode->getBasePtr(),
Base, Scale, Index, Disp, Segment))
break;
case InlineAsm::Constraint_v: // not offsetable ??
case InlineAsm::Constraint_m: // memory
case InlineAsm::Constraint_X:
- if (!SelectAddr(nullptr, Op, Op0, Op1, Op2, Op3, Op4))
+ if (!selectAddr(nullptr, Op, Op0, Op1, Op2, Op3, Op4))
return true;
break;
}
return false;
}
-/// createX86ISelDag - This pass converts a legalized DAG into a
-/// X86-specific DAG, ready for instruction scheduling.
-///
+/// This pass converts a legalized DAG into a X86-specific DAG,
+/// ready for instruction scheduling.
FunctionPass *llvm::createX86ISelDag(X86TargetMachine &TM,
CodeGenOpt::Level OptLevel) {
return new X86DAGToDAGISel(TM, OptLevel);