#include "llvm/Intrinsics.h"
#include "llvm/Support/CFG.h"
#include "llvm/Type.h"
+#include "llvm/CodeGen/FunctionLoweringInfo.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
virtual void PreprocessISelDAG();
+ inline bool immSext8(SDNode *N) const {
+ return isInt<8>(cast<ConstantSDNode>(N)->getSExtValue());
+ }
+
+ // i64immSExt32 predicate - 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;
+ }
+
// Include the pieces autogenerated from the target description.
#include "X86GenDAGISel.inc"
SDNode *Select(SDNode *N);
SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
SDNode *SelectAtomicLoadAdd(SDNode *Node, EVT NVT);
+ SDNode *SelectAtomicLoadArith(SDNode *Node, EVT NVT);
- bool MatchSegmentBaseAddress(SDValue N, X86ISelAddressMode &AM);
- bool MatchLoad(SDValue N, X86ISelAddressMode &AM);
+ 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,
unsigned Depth);
bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
- bool SelectAddr(SDNode *Op, SDValue N, SDValue &Base,
+ bool SelectAddr(SDNode *Parent, SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index, SDValue &Disp,
SDValue &Segment);
- bool SelectLEAAddr(SDNode *Op, SDValue N, SDValue &Base,
- SDValue &Scale, SDValue &Index, SDValue &Disp);
- bool SelectTLSADDRAddr(SDNode *Op, SDValue N, SDValue &Base,
- SDValue &Scale, SDValue &Index, SDValue &Disp);
+ bool SelectLEAAddr(SDValue N, SDValue &Base,
+ SDValue &Scale, SDValue &Index, SDValue &Disp,
+ SDValue &Segment);
+ bool SelectTLSADDRAddr(SDValue N, SDValue &Base,
+ SDValue &Scale, SDValue &Index, SDValue &Disp,
+ SDValue &Segment);
bool SelectScalarSSELoad(SDNode *Root, SDValue N,
SDValue &Base, SDValue &Scale,
SDValue &Index, SDValue &Disp,
// These are 32-bit even in 64-bit mode since RIP relative offset
// is 32-bit.
if (AM.GV)
- Disp = CurDAG->getTargetGlobalAddress(AM.GV, MVT::i32, AM.Disp,
+ Disp = CurDAG->getTargetGlobalAddress(AM.GV, DebugLoc(),
+ MVT::i32, AM.Disp,
AM.SymbolFlags);
else if (AM.CP)
Disp = CurDAG->getTargetConstantPool(AM.CP, MVT::i32,
return CurDAG->getTargetConstant(Imm, MVT::i8);
}
- /// getI16Imm - Return a target constant with the specified value, of type
- /// i16.
- inline SDValue getI16Imm(unsigned Imm) {
- return CurDAG->getTargetConstant(Imm, MVT::i16);
- }
-
/// getI32Imm - Return a target constant with the specified value, of type
/// i32.
inline SDValue getI32Imm(unsigned Imm) {
if (N->getOpcode() != ISD::FP_ROUND && N->getOpcode() != ISD::FP_EXTEND)
continue;
- // If the source and destination are SSE registers, then this is a legal
- // conversion that should not be lowered.
EVT SrcVT = N->getOperand(0).getValueType();
EVT DstVT = N->getValueType(0);
+
+ // If any of the sources are vectors, no fp stack involved.
+ if (SrcVT.isVector() || DstVT.isVector())
+ continue;
+
+ // If the source and destination are SSE registers, then this is a legal
+ // conversion that should not be lowered.
bool SrcIsSSE = X86Lowering.isScalarFPTypeInSSEReg(SrcVT);
bool DstIsSSE = X86Lowering.isScalarFPTypeInSSEReg(DstVT);
if (SrcIsSSE && DstIsSSE)
// FIXME: optimize the case where the src/dest is a load or store?
SDValue Store = CurDAG->getTruncStore(CurDAG->getEntryNode(), dl,
N->getOperand(0),
- MemTmp, NULL, 0, MemVT,
+ MemTmp, MachinePointerInfo(), MemVT,
false, false, 0);
SDValue Result = CurDAG->getExtLoad(ISD::EXTLOAD, dl, DstVT, Store, MemTmp,
- NULL, 0, MemVT, false, false, 0);
+ MachinePointerInfo(),
+ MemVT, false, false, 0);
// We're about to replace all uses of the FP_ROUND/FP_EXTEND with the
// extload we created. This will cause general havok on the dag because
void X86DAGToDAGISel::EmitSpecialCodeForMain(MachineBasicBlock *BB,
MachineFrameInfo *MFI) {
const TargetInstrInfo *TII = TM.getInstrInfo();
- if (Subtarget->isTargetCygMing())
+ if (Subtarget->isTargetCygMing()) {
+ unsigned CallOp =
+ Subtarget->is64Bit() ? X86::WINCALL64pcrel32 : X86::CALLpcrel32;
BuildMI(BB, DebugLoc(),
- TII->get(X86::CALLpcrel32)).addExternalSymbol("__main");
+ TII->get(CallOp)).addExternalSymbol("__main");
+ }
}
void X86DAGToDAGISel::EmitFunctionEntryCode() {
EmitSpecialCodeForMain(MF->begin(), MF->getFrameInfo());
}
+static bool isDispSafeForFrameIndex(int64_t Val) {
+ // On 64-bit platforms, we can run into an issue where a frame index
+ // includes a displacement that, when added to the explicit displacement,
+ // will overflow the displacement field. Assuming that the frame index
+ // displacement fits into a 31-bit integer (which is only slightly more
+ // aggressive than the current fundamental assumption that it fits into
+ // a 32-bit integer), a 31-bit disp should always be safe.
+ return isInt<31>(Val);
+}
-bool X86DAGToDAGISel::MatchSegmentBaseAddress(SDValue N,
- X86ISelAddressMode &AM) {
- assert(N.getOpcode() == X86ISD::SegmentBaseAddress);
- SDValue Segment = N.getOperand(0);
-
- if (AM.Segment.getNode() == 0) {
- AM.Segment = Segment;
- return false;
+bool X86DAGToDAGISel::FoldOffsetIntoAddress(uint64_t Offset,
+ X86ISelAddressMode &AM) {
+ int64_t Val = AM.Disp + Offset;
+ CodeModel::Model M = TM.getCodeModel();
+ if (Subtarget->is64Bit()) {
+ if (!X86::isOffsetSuitableForCodeModel(Val, M,
+ AM.hasSymbolicDisplacement()))
+ return true;
+ // In addition to the checks required for a register base, check that
+ // we do not try to use an unsafe Disp with a frame index.
+ if (AM.BaseType == X86ISelAddressMode::FrameIndexBase &&
+ !isDispSafeForFrameIndex(Val))
+ return true;
}
+ AM.Disp = Val;
+ return false;
- return true;
}
-bool X86DAGToDAGISel::MatchLoad(SDValue N, X86ISelAddressMode &AM) {
+bool X86DAGToDAGISel::MatchLoadInAddress(LoadSDNode *N, X86ISelAddressMode &AM){
+ SDValue Address = N->getOperand(1);
+
+ // load gs:0 -> GS segment register.
+ // load fs:0 -> FS segment register.
+ //
// This optimization is valid because the GNU TLS model defines that
// gs:0 (or fs:0 on X86-64) contains its own address.
// For more information see http://people.redhat.com/drepper/tls.pdf
-
- SDValue Address = N.getOperand(1);
- if (Address.getOpcode() == X86ISD::SegmentBaseAddress &&
- !MatchSegmentBaseAddress (Address, AM))
- return false;
-
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Address))
+ if (C->getSExtValue() == 0 && AM.Segment.getNode() == 0 &&
+ Subtarget->isTargetELF())
+ switch (N->getPointerInfo().getAddrSpace()) {
+ case 256:
+ AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
+ return false;
+ case 257:
+ AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
+ return false;
+ }
+
return true;
}
// must allow RIP.
!AM.hasBaseOrIndexReg() && N.getOpcode() == X86ISD::WrapperRIP) {
if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(N0)) {
- int64_t Offset = AM.Disp + G->getOffset();
- if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
+ X86ISelAddressMode Backup = AM;
AM.GV = G->getGlobal();
- AM.Disp = Offset;
AM.SymbolFlags = G->getTargetFlags();
+ if (FoldOffsetIntoAddress(G->getOffset(), AM)) {
+ AM = Backup;
+ return true;
+ }
} else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(N0)) {
- int64_t Offset = AM.Disp + CP->getOffset();
- if (!X86::isOffsetSuitableForCodeModel(Offset, M)) return true;
+ X86ISelAddressMode Backup = AM;
AM.CP = CP->getConstVal();
AM.Align = CP->getAlignment();
- AM.Disp = Offset;
AM.SymbolFlags = CP->getTargetFlags();
+ if (FoldOffsetIntoAddress(CP->getOffset(), AM)) {
+ AM = Backup;
+ return true;
+ }
} else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(N0)) {
AM.ES = S->getSymbol();
AM.SymbolFlags = S->getTargetFlags();
return false;
}
-/// isLogicallyAddWithConstant - Return true if this node is semantically an
-/// add of a value with a constantint.
-static bool isLogicallyAddWithConstant(SDValue V, SelectionDAG *CurDAG) {
- // Check for (add x, Cst)
- if (V->getOpcode() == ISD::ADD)
- return isa<ConstantSDNode>(V->getOperand(1));
+// 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) {
+ if (N.getNode()->getNodeId() == -1 ||
+ N.getNode()->getNodeId() > Pos.getNode()->getNodeId()) {
+ DAG.RepositionNode(Pos.getNode(), N.getNode());
+ N.getNode()->setNodeId(Pos.getNode()->getNodeId());
+ }
+}
- // Check for (or x, Cst), where Cst & x == 0.
- if (V->getOpcode() != ISD::OR ||
- !isa<ConstantSDNode>(V->getOperand(1)))
- return false;
-
- // Handle "X | C" as "X + C" iff X is known to have C bits clear.
- ConstantSDNode *CN = cast<ConstantSDNode>(V->getOperand(1));
-
- // Check to see if the LHS & C is zero.
- return CurDAG->MaskedValueIsZero(V->getOperand(0), CN->getAPIntValue());
+// Transform "(X >> (8-C1)) & C2" to "(X >> 8) & 0xff)" if 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,
+ uint64_t Mask,
+ SDValue Shift, SDValue X,
+ X86ISelAddressMode &AM) {
+ if (Shift.getOpcode() != ISD::SRL ||
+ !isa<ConstantSDNode>(Shift.getOperand(1)) ||
+ !Shift.hasOneUse())
+ return true;
+
+ int ScaleLog = 8 - Shift.getConstantOperandVal(1);
+ if (ScaleLog <= 0 || ScaleLog >= 4 ||
+ Mask != (0xffu << ScaleLog))
+ return true;
+
+ EVT VT = N.getValueType();
+ DebugLoc DL = N.getDebugLoc();
+ SDValue Eight = DAG.getConstant(8, MVT::i8);
+ SDValue NewMask = DAG.getConstant(0xff, VT);
+ SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, X, Eight);
+ SDValue And = DAG.getNode(ISD::AND, DL, VT, Srl, NewMask);
+ SDValue ShlCount = DAG.getConstant(ScaleLog, MVT::i8);
+ SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, And, ShlCount);
+
+ // Insert the new nodes into the topological ordering. We must do this in
+ // a valid topological ordering as nothing is going to go back and re-sort
+ // 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);
+ DAG.ReplaceAllUsesWith(N, Shl);
+ AM.IndexReg = And;
+ AM.Scale = (1 << ScaleLog);
+ return false;
+}
+
+// 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,
+ uint64_t Mask,
+ SDValue Shift, SDValue X,
+ X86ISelAddressMode &AM) {
+ if (Shift.getOpcode() != ISD::SHL ||
+ !isa<ConstantSDNode>(Shift.getOperand(1)))
+ return true;
+
+ // Not likely to be profitable if either the AND or SHIFT node has more
+ // than one use (unless all uses are for address computation). Besides,
+ // isel mechanism requires their node ids to be reused.
+ if (!N.hasOneUse() || !Shift.hasOneUse())
+ return true;
+
+ // Verify that the shift amount is something we can fold.
+ unsigned ShiftAmt = Shift.getConstantOperandVal(1);
+ if (ShiftAmt != 1 && ShiftAmt != 2 && ShiftAmt != 3)
+ return true;
+
+ EVT VT = N.getValueType();
+ DebugLoc DL = N.getDebugLoc();
+ SDValue NewMask = DAG.getConstant(Mask >> ShiftAmt, VT);
+ SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, NewMask);
+ SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, NewAnd, Shift.getOperand(1));
+
+ // Insert the new nodes into the topological ordering. We must do this in
+ // a valid topological ordering as nothing is going to go back and re-sort
+ // 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);
+ DAG.ReplaceAllUsesWith(N, NewShift);
+
+ AM.Scale = 1 << ShiftAmt;
+ AM.IndexReg = NewAnd;
+ return false;
+}
+
+// Implement some heroics to detect shifts of masked values where the mask can
+// be replaced by extending the shift and undoing that in the addressing mode
+// scale. Patterns such as (shl (srl x, c1), c2) are canonicalized into (and
+// (srl x, SHIFT), MASK) by DAGCombines that don't know the shl can be done in
+// the addressing mode. This results in code such as:
+//
+// int f(short *y, int *lookup_table) {
+// ...
+// return *y + lookup_table[*y >> 11];
+// }
+//
+// Turning into:
+// movzwl (%rdi), %eax
+// movl %eax, %ecx
+// shrl $11, %ecx
+// addl (%rsi,%rcx,4), %eax
+//
+// Instead of:
+// movzwl (%rdi), %eax
+// movl %eax, %ecx
+// shrl $9, %ecx
+// andl $124, %rcx
+// addl (%rsi,%rcx), %eax
+//
+// 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,
+ uint64_t Mask,
+ SDValue Shift, SDValue X,
+ X86ISelAddressMode &AM) {
+ if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse() ||
+ !isa<ConstantSDNode>(Shift.getOperand(1)))
+ return true;
+
+ unsigned ShiftAmt = Shift.getConstantOperandVal(1);
+ unsigned MaskLZ = CountLeadingZeros_64(Mask);
+ unsigned MaskTZ = CountTrailingZeros_64(Mask);
+
+ // The amount of shift we're trying to fit into the addressing mode is taken
+ // from the trailing zeros of the mask.
+ unsigned AMShiftAmt = MaskTZ;
+
+ // There is nothing we can do here unless the mask is removing some bits.
+ // Also, the addressing mode can only represent shifts of 1, 2, or 3 bits.
+ if (AMShiftAmt <= 0 || AMShiftAmt > 3) return true;
+
+ // We also need to ensure that mask is a continuous run of bits.
+ if (CountTrailingOnes_64(Mask >> MaskTZ) + MaskTZ + MaskLZ != 64) return true;
+
+ // Scale the leading zero count down based on the actual size of the value.
+ // Also scale it down based on the size of the shift.
+ MaskLZ -= (64 - X.getValueSizeInBits()) + ShiftAmt;
+
+ // The final check is to ensure that any masked out high bits of X are
+ // already known to be zero. Otherwise, the mask has a semantic impact
+ // other than masking out a couple of low bits. Unfortunately, because of
+ // the mask, zero extensions will be removed from operands in some cases.
+ // This code works extra hard to look through extensions because we can
+ // replace them with zero extensions cheaply if necessary.
+ bool ReplacingAnyExtend = false;
+ if (X.getOpcode() == ISD::ANY_EXTEND) {
+ unsigned ExtendBits =
+ X.getValueSizeInBits() - X.getOperand(0).getValueSizeInBits();
+ // Assume that we'll replace the any-extend with a zero-extend, and
+ // narrow the search to the extended value.
+ X = X.getOperand(0);
+ MaskLZ = ExtendBits > MaskLZ ? 0 : MaskLZ - ExtendBits;
+ ReplacingAnyExtend = true;
+ }
+ APInt MaskedHighBits = APInt::getHighBitsSet(X.getValueSizeInBits(),
+ MaskLZ);
+ APInt KnownZero, KnownOne;
+ DAG.ComputeMaskedBits(X, MaskedHighBits, KnownZero, KnownOne);
+ if (MaskedHighBits != KnownZero) return true;
+
+ // We've identified a pattern that can be transformed into a single shift
+ // and an addressing mode. Make it so.
+ EVT VT = N.getValueType();
+ if (ReplacingAnyExtend) {
+ assert(X.getValueType() != VT);
+ // We looked through an ANY_EXTEND node, insert a ZERO_EXTEND.
+ SDValue NewX = DAG.getNode(ISD::ZERO_EXTEND, X.getDebugLoc(), VT, X);
+ InsertDAGNode(DAG, N, NewX);
+ X = NewX;
+ }
+ DebugLoc DL = N.getDebugLoc();
+ SDValue NewSRLAmt = DAG.getConstant(ShiftAmt + AMShiftAmt, MVT::i8);
+ SDValue NewSRL = DAG.getNode(ISD::SRL, DL, VT, X, NewSRLAmt);
+ SDValue NewSHLAmt = DAG.getConstant(AMShiftAmt, MVT::i8);
+ SDValue NewSHL = DAG.getNode(ISD::SHL, DL, VT, NewSRL, NewSHLAmt);
+
+ // Insert the new nodes into the topological ordering. We must do this in
+ // a valid topological ordering as nothing is going to go back and re-sort
+ // 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);
+ DAG.ReplaceAllUsesWith(N, NewSHL);
+
+ AM.Scale = 1 << AMShiftAmt;
+ AM.IndexReg = NewSRL;
+ return false;
}
bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
unsigned Depth) {
- bool is64Bit = Subtarget->is64Bit();
DebugLoc dl = N.getDebugLoc();
DEBUG({
dbgs() << "MatchAddress: ";
if (Depth > 5)
return MatchAddressBase(N, AM);
- CodeModel::Model M = TM.getCodeModel();
-
// 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.
// RIP relative addressing: %rip + 32-bit displacement!
// consistency.
if (!AM.ES && AM.JT != -1) return true;
- if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N)) {
- int64_t Val = AM.Disp + Cst->getSExtValue();
- if (X86::isOffsetSuitableForCodeModel(Val, M,
- AM.hasSymbolicDisplacement())) {
- AM.Disp = Val;
+ if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N))
+ if (!FoldOffsetIntoAddress(Cst->getSExtValue(), AM))
return false;
- }
- }
return true;
}
default: break;
case ISD::Constant: {
uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
- if (!is64Bit ||
- X86::isOffsetSuitableForCodeModel(AM.Disp + Val, M,
- AM.hasSymbolicDisplacement())) {
- AM.Disp += Val;
+ if (!FoldOffsetIntoAddress(Val, AM))
return false;
- }
break;
}
- case X86ISD::SegmentBaseAddress:
- if (!MatchSegmentBaseAddress(N, AM))
- return false;
- break;
-
case X86ISD::Wrapper:
case X86ISD::WrapperRIP:
if (!MatchWrapper(N, AM))
break;
case ISD::LOAD:
- if (!MatchLoad(N, AM))
+ if (!MatchLoadInAddress(cast<LoadSDNode>(N), AM))
return false;
break;
case ISD::FrameIndex:
- if (AM.BaseType == X86ISelAddressMode::RegBase
- && AM.Base_Reg.getNode() == 0) {
+ if (AM.BaseType == X86ISelAddressMode::RegBase &&
+ AM.Base_Reg.getNode() == 0 &&
+ (!Subtarget->is64Bit() || isDispSafeForFrameIndex(AM.Disp))) {
AM.BaseType = X86ISelAddressMode::FrameIndexBase;
AM.Base_FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
return false;
// Okay, we know that we have a scale by now. However, if the scaled
// value is an add of something and a constant, we can fold the
// constant into the disp field here.
- if (isLogicallyAddWithConstant(ShVal, CurDAG)) {
+ if (CurDAG->isBaseWithConstantOffset(ShVal)) {
AM.IndexReg = ShVal.getNode()->getOperand(0);
ConstantSDNode *AddVal =
cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
- uint64_t Disp = AM.Disp + (AddVal->getSExtValue() << Val);
- if (!is64Bit ||
- X86::isOffsetSuitableForCodeModel(Disp, M,
- AM.hasSymbolicDisplacement()))
- AM.Disp = Disp;
- else
- AM.IndexReg = ShVal;
- } else {
- AM.IndexReg = ShVal;
+ uint64_t Disp = AddVal->getSExtValue() << Val;
+ if (!FoldOffsetIntoAddress(Disp, AM))
+ return false;
}
+
+ AM.IndexReg = ShVal;
return false;
}
break;
}
+ case ISD::SRL: {
+ // Scale must not be used already.
+ if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
+
+ SDValue And = N.getOperand(0);
+ if (And.getOpcode() != ISD::AND) break;
+ SDValue X = And.getOperand(0);
+
+ // We only handle up to 64-bit values here as those are what matter for
+ // addressing mode optimizations.
+ if (X.getValueSizeInBits() > 64) break;
+
+ // The mask used for the transform is expected to be post-shift, but we
+ // found the shift first so just apply the shift to the mask before passing
+ // it down.
+ if (!isa<ConstantSDNode>(N.getOperand(1)) ||
+ !isa<ConstantSDNode>(And.getOperand(1)))
+ break;
+ uint64_t Mask = And.getConstantOperandVal(1) >> N.getConstantOperandVal(1);
+
+ // Try to fold the mask and shift into the scale, and return false if we
+ // succeed.
+ if (!FoldMaskAndShiftToScale(*CurDAG, N, Mask, N, X, AM))
+ return false;
+ break;
+ }
+
case ISD::SMUL_LOHI:
case ISD::UMUL_LOHI:
// A mul_lohi where we need the low part can be folded as a plain multiply.
Reg = MulVal.getNode()->getOperand(0);
ConstantSDNode *AddVal =
cast<ConstantSDNode>(MulVal.getNode()->getOperand(1));
- uint64_t Disp = AM.Disp + AddVal->getSExtValue() *
- CN->getZExtValue();
- if (!is64Bit ||
- X86::isOffsetSuitableForCodeModel(Disp, M,
- AM.hasSymbolicDisplacement()))
- AM.Disp = Disp;
- else
+ uint64_t Disp = AddVal->getSExtValue() * CN->getZExtValue();
+ if (FoldOffsetIntoAddress(Disp, AM))
Reg = N.getNode()->getOperand(0);
} else {
Reg = N.getNode()->getOperand(0);
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());
- }
+ InsertDAGNode(*CurDAG, N, Zero);
+ InsertDAGNode(*CurDAG, N, Neg);
return false;
}
// 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);
- SDValue LHS = Handle.getValue().getNode()->getOperand(0);
- SDValue RHS = Handle.getValue().getNode()->getOperand(1);
X86ISelAddressMode Backup = AM;
- if (!MatchAddressRecursively(LHS, AM, Depth+1) &&
- !MatchAddressRecursively(RHS, AM, Depth+1))
+ if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
+ !MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1))
return false;
AM = Backup;
- LHS = Handle.getValue().getNode()->getOperand(0);
- RHS = Handle.getValue().getNode()->getOperand(1);
-
+
// Try again after commuting the operands.
- if (!MatchAddressRecursively(RHS, AM, Depth+1) &&
- !MatchAddressRecursively(LHS, AM, Depth+1))
+ if (!MatchAddressRecursively(Handle.getValue().getOperand(1), AM, Depth+1)&&
+ !MatchAddressRecursively(Handle.getValue().getOperand(0), AM, Depth+1))
return false;
AM = Backup;
- LHS = Handle.getValue().getNode()->getOperand(0);
- RHS = Handle.getValue().getNode()->getOperand(1);
// 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
if (AM.BaseType == X86ISelAddressMode::RegBase &&
!AM.Base_Reg.getNode() &&
!AM.IndexReg.getNode()) {
- AM.Base_Reg = LHS;
- AM.IndexReg = RHS;
+ 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 (isLogicallyAddWithConstant(N, CurDAG)) {
+ if (CurDAG->isBaseWithConstantOffset(N)) {
X86ISelAddressMode Backup = AM;
ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
- uint64_t Offset = CN->getSExtValue();
// Start with the LHS as an addr mode.
if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
- // Address could not have picked a GV address for the displacement.
- AM.GV == NULL &&
- // On x86-64, the resultant disp must fit in 32-bits.
- (!is64Bit ||
- X86::isOffsetSuitableForCodeModel(AM.Disp + Offset, M,
- AM.hasSymbolicDisplacement()))) {
- AM.Disp += Offset;
+ !FoldOffsetIntoAddress(CN->getSExtValue(), AM))
return false;
- }
AM = Backup;
}
break;
// Perform some heroic transforms on an and of a constant-count shift
// with a constant to enable use of the scaled offset field.
- SDValue Shift = N.getOperand(0);
- if (Shift.getNumOperands() != 2) break;
-
// Scale must not be used already.
if (AM.IndexReg.getNode() != 0 || AM.Scale != 1) break;
+ SDValue Shift = N.getOperand(0);
+ if (Shift.getOpcode() != ISD::SRL && Shift.getOpcode() != ISD::SHL) break;
SDValue X = Shift.getOperand(0);
- ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N.getOperand(1));
- ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
- if (!C1 || !C2) break;
-
- // Handle "(X >> (8-C1)) & C2" as "(X >> 8) & 0xff)" if safe. This
- // allows us to convert the shift and and into an h-register extract and
- // a scaled index.
- if (Shift.getOpcode() == ISD::SRL && Shift.hasOneUse()) {
- unsigned ScaleLog = 8 - C1->getZExtValue();
- if (ScaleLog > 0 && ScaleLog < 4 &&
- C2->getZExtValue() == (UINT64_C(0xff) << ScaleLog)) {
- SDValue Eight = CurDAG->getConstant(8, MVT::i8);
- SDValue Mask = CurDAG->getConstant(0xff, N.getValueType());
- SDValue Srl = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
- X, Eight);
- SDValue And = CurDAG->getNode(ISD::AND, dl, N.getValueType(),
- Srl, Mask);
- SDValue ShlCount = CurDAG->getConstant(ScaleLog, MVT::i8);
- SDValue Shl = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
- And, ShlCount);
-
- // Insert the new nodes into the topological ordering.
- if (Eight.getNode()->getNodeId() == -1 ||
- Eight.getNode()->getNodeId() > X.getNode()->getNodeId()) {
- CurDAG->RepositionNode(X.getNode(), Eight.getNode());
- Eight.getNode()->setNodeId(X.getNode()->getNodeId());
- }
- if (Mask.getNode()->getNodeId() == -1 ||
- Mask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
- CurDAG->RepositionNode(X.getNode(), Mask.getNode());
- Mask.getNode()->setNodeId(X.getNode()->getNodeId());
- }
- if (Srl.getNode()->getNodeId() == -1 ||
- Srl.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
- CurDAG->RepositionNode(Shift.getNode(), Srl.getNode());
- Srl.getNode()->setNodeId(Shift.getNode()->getNodeId());
- }
- if (And.getNode()->getNodeId() == -1 ||
- And.getNode()->getNodeId() > N.getNode()->getNodeId()) {
- CurDAG->RepositionNode(N.getNode(), And.getNode());
- And.getNode()->setNodeId(N.getNode()->getNodeId());
- }
- if (ShlCount.getNode()->getNodeId() == -1 ||
- ShlCount.getNode()->getNodeId() > X.getNode()->getNodeId()) {
- CurDAG->RepositionNode(X.getNode(), ShlCount.getNode());
- ShlCount.getNode()->setNodeId(N.getNode()->getNodeId());
- }
- if (Shl.getNode()->getNodeId() == -1 ||
- Shl.getNode()->getNodeId() > N.getNode()->getNodeId()) {
- CurDAG->RepositionNode(N.getNode(), Shl.getNode());
- Shl.getNode()->setNodeId(N.getNode()->getNodeId());
- }
- CurDAG->ReplaceAllUsesWith(N, Shl);
- AM.IndexReg = And;
- AM.Scale = (1 << ScaleLog);
- return false;
- }
- }
- // Handle "(X << C1) & C2" as "(X & (C2>>C1)) << C1" if safe and if this
- // allows us to fold the shift into this addressing mode.
- if (Shift.getOpcode() != ISD::SHL) break;
+ // We only handle up to 64-bit values here as those are what matter for
+ // addressing mode optimizations.
+ if (X.getValueSizeInBits() > 64) break;
- // Not likely to be profitable if either the AND or SHIFT node has more
- // than one use (unless all uses are for address computation). Besides,
- // isel mechanism requires their node ids to be reused.
- if (!N.hasOneUse() || !Shift.hasOneUse())
+ if (!isa<ConstantSDNode>(N.getOperand(1)))
break;
-
- // Verify that the shift amount is something we can fold.
- unsigned ShiftCst = C1->getZExtValue();
- if (ShiftCst != 1 && ShiftCst != 2 && ShiftCst != 3)
- break;
-
- // Get the new AND mask, this folds to a constant.
- SDValue NewANDMask = CurDAG->getNode(ISD::SRL, dl, N.getValueType(),
- SDValue(C2, 0), SDValue(C1, 0));
- SDValue NewAND = CurDAG->getNode(ISD::AND, dl, N.getValueType(), X,
- NewANDMask);
- SDValue NewSHIFT = CurDAG->getNode(ISD::SHL, dl, N.getValueType(),
- NewAND, SDValue(C1, 0));
+ uint64_t Mask = N.getConstantOperandVal(1);
- // Insert the new nodes into the topological ordering.
- if (C1->getNodeId() > X.getNode()->getNodeId()) {
- CurDAG->RepositionNode(X.getNode(), C1);
- C1->setNodeId(X.getNode()->getNodeId());
- }
- if (NewANDMask.getNode()->getNodeId() == -1 ||
- NewANDMask.getNode()->getNodeId() > X.getNode()->getNodeId()) {
- CurDAG->RepositionNode(X.getNode(), NewANDMask.getNode());
- NewANDMask.getNode()->setNodeId(X.getNode()->getNodeId());
- }
- if (NewAND.getNode()->getNodeId() == -1 ||
- NewAND.getNode()->getNodeId() > Shift.getNode()->getNodeId()) {
- CurDAG->RepositionNode(Shift.getNode(), NewAND.getNode());
- NewAND.getNode()->setNodeId(Shift.getNode()->getNodeId());
- }
- if (NewSHIFT.getNode()->getNodeId() == -1 ||
- NewSHIFT.getNode()->getNodeId() > N.getNode()->getNodeId()) {
- CurDAG->RepositionNode(N.getNode(), NewSHIFT.getNode());
- NewSHIFT.getNode()->setNodeId(N.getNode()->getNodeId());
- }
+ // Try to fold the mask and shift into an extract and scale.
+ if (!FoldMaskAndShiftToExtract(*CurDAG, N, Mask, Shift, X, AM))
+ return false;
- CurDAG->ReplaceAllUsesWith(N, NewSHIFT);
-
- AM.Scale = 1 << ShiftCst;
- AM.IndexReg = NewAND;
- return false;
+ // Try to fold the mask and shift directly into the scale.
+ 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))
+ return false;
+ break;
}
}
/// SelectAddr - returns true if it is able pattern match an addressing mode.
/// It returns the operands which make up the maximal addressing mode it can
/// match by reference.
-bool X86DAGToDAGISel::SelectAddr(SDNode *Op, SDValue N, SDValue &Base,
+///
+/// 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,
SDValue &Scale, SDValue &Index,
SDValue &Disp, SDValue &Segment) {
X86ISelAddressMode AM;
+
+ if (Parent &&
+ // This list of opcodes are all the nodes that have an "addr:$ptr" operand
+ // that are not a MemSDNode, and thus don't have proper addrspace info.
+ Parent->getOpcode() != ISD::INTRINSIC_W_CHAIN && // unaligned loads, fixme
+ Parent->getOpcode() != ISD::INTRINSIC_VOID && // nontemporal stores
+ Parent->getOpcode() != X86ISD::TLSCALL) { // Fixme
+ unsigned AddrSpace =
+ cast<MemSDNode>(Parent)->getPointerInfo().getAddrSpace();
+ // AddrSpace 256 -> GS, 257 -> FS.
+ if (AddrSpace == 256)
+ AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
+ if (AddrSpace == 257)
+ AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
+ }
+
if (MatchAddress(N, AM))
return false;
IsProfitableToFold(N.getOperand(0), N.getNode(), Root) &&
IsLegalToFold(N.getOperand(0), N.getNode(), Root, OptLevel)) {
LoadSDNode *LD = cast<LoadSDNode>(PatternNodeWithChain);
- if (!SelectAddr(Root, 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(Root, 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;
/// SelectLEAAddr - it calls SelectAddr and determines if the maximal addressing
/// mode it matches can be cost effectively emitted as an LEA instruction.
-bool X86DAGToDAGISel::SelectLEAAddr(SDNode *Op, SDValue N,
+bool X86DAGToDAGISel::SelectLEAAddr(SDValue N,
SDValue &Base, SDValue &Scale,
- SDValue &Index, SDValue &Disp) {
+ SDValue &Index, SDValue &Disp,
+ SDValue &Segment) {
X86ISelAddressMode AM;
// Set AM.Segment to prevent MatchAddress from using one. LEA doesn't support
if (Complexity <= 2)
return false;
- SDValue Segment;
getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
return true;
}
/// SelectTLSADDRAddr - This is only run on TargetGlobalTLSAddress nodes.
-bool X86DAGToDAGISel::SelectTLSADDRAddr(SDNode *Op, SDValue N, SDValue &Base,
+bool X86DAGToDAGISel::SelectTLSADDRAddr(SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index,
- SDValue &Disp) {
+ SDValue &Disp, SDValue &Segment) {
assert(N.getOpcode() == ISD::TargetGlobalTLSAddress);
const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N);
AM.IndexReg = CurDAG->getRegister(0, MVT::i64);
}
- SDValue Segment;
getAddressOperands(AM, Base, Scale, Index, Disp, Segment);
return true;
}
!IsLegalToFold(N, P, P, OptLevel))
return false;
- return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp, Segment);
+ return SelectAddr(N.getNode(),
+ N.getOperand(1), Base, Scale, Index, Disp, Segment);
}
/// getGlobalBaseReg - Return an SDNode that returns the value of
return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
}
-static SDNode *FindCallStartFromCall(SDNode *Node) {
- if (Node->getOpcode() == ISD::CALLSEQ_START) return Node;
- assert(Node->getOperand(0).getValueType() == MVT::Other &&
- "Node doesn't have a token chain argument!");
- return FindCallStartFromCall(Node->getOperand(0).getNode());
-}
-
SDNode *X86DAGToDAGISel::SelectAtomic64(SDNode *Node, unsigned Opc) {
SDValue Chain = Node->getOperand(0);
SDValue In1 = Node->getOperand(1);
SDValue In2L = Node->getOperand(2);
SDValue In2H = Node->getOperand(3);
SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
- if (!SelectAddr(In1.getNode(), In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
+ if (!SelectAddr(Node, In1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
return NULL;
MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
return ResNode;
}
+// FIXME: Figure out some way to unify this with the 'or' and other code
+// below.
SDNode *X86DAGToDAGISel::SelectAtomicLoadAdd(SDNode *Node, EVT NVT) {
if (Node->hasAnyUseOfValue(0))
return 0;
SDValue Ptr = Node->getOperand(1);
SDValue Val = Node->getOperand(2);
SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
- if (!SelectAddr(Ptr.getNode(), Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
+ if (!SelectAddr(Node, Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
return 0;
bool isInc = false, isDec = false, isSub = false, isCN = false;
ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val);
- if (CN) {
+ if (CN && CN->getSExtValue() == (int32_t)CN->getSExtValue()) {
isCN = true;
int64_t CNVal = CN->getSExtValue();
if (CNVal == 1)
Val = Val.getOperand(1);
}
+ DebugLoc dl = Node->getDebugLoc();
unsigned Opc = 0;
switch (NVT.getSimpleVT().SimpleTy) {
default: return 0;
Opc = X86::LOCK_DEC16m;
else if (isSub) {
if (isCN) {
- if (Predicate_immSext8(Val.getNode()))
+ if (immSext8(Val.getNode()))
Opc = X86::LOCK_SUB16mi8;
else
Opc = X86::LOCK_SUB16mi;
Opc = X86::LOCK_SUB16mr;
} else {
if (isCN) {
- if (Predicate_immSext8(Val.getNode()))
+ if (immSext8(Val.getNode()))
Opc = X86::LOCK_ADD16mi8;
else
Opc = X86::LOCK_ADD16mi;
Opc = X86::LOCK_DEC32m;
else if (isSub) {
if (isCN) {
- if (Predicate_immSext8(Val.getNode()))
+ if (immSext8(Val.getNode()))
Opc = X86::LOCK_SUB32mi8;
else
Opc = X86::LOCK_SUB32mi;
Opc = X86::LOCK_SUB32mr;
} else {
if (isCN) {
- if (Predicate_immSext8(Val.getNode()))
+ if (immSext8(Val.getNode()))
Opc = X86::LOCK_ADD32mi8;
else
Opc = X86::LOCK_ADD32mi;
else if (isSub) {
Opc = X86::LOCK_SUB64mr;
if (isCN) {
- if (Predicate_immSext8(Val.getNode()))
+ if (immSext8(Val.getNode()))
Opc = X86::LOCK_SUB64mi8;
- else if (Predicate_i64immSExt32(Val.getNode()))
+ else if (i64immSExt32(Val.getNode()))
Opc = X86::LOCK_SUB64mi32;
}
} else {
Opc = X86::LOCK_ADD64mr;
if (isCN) {
- if (Predicate_immSext8(Val.getNode()))
+ if (immSext8(Val.getNode()))
Opc = X86::LOCK_ADD64mi8;
- else if (Predicate_i64immSExt32(Val.getNode()))
+ else if (i64immSExt32(Val.getNode()))
Opc = X86::LOCK_ADD64mi32;
}
}
break;
}
- DebugLoc dl = Node->getDebugLoc();
SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
dl, NVT), 0);
MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
}
}
+enum AtomicOpc {
+ OR,
+ AND,
+ XOR,
+ AtomicOpcEnd
+};
+
+enum AtomicSz {
+ ConstantI8,
+ I8,
+ SextConstantI16,
+ ConstantI16,
+ I16,
+ SextConstantI32,
+ ConstantI32,
+ I32,
+ SextConstantI64,
+ ConstantI64,
+ I64,
+ AtomicSzEnd
+};
+
+static const unsigned int AtomicOpcTbl[AtomicOpcEnd][AtomicSzEnd] = {
+ {
+ X86::LOCK_OR8mi,
+ X86::LOCK_OR8mr,
+ X86::LOCK_OR16mi8,
+ X86::LOCK_OR16mi,
+ X86::LOCK_OR16mr,
+ X86::LOCK_OR32mi8,
+ X86::LOCK_OR32mi,
+ X86::LOCK_OR32mr,
+ X86::LOCK_OR64mi8,
+ X86::LOCK_OR64mi32,
+ X86::LOCK_OR64mr
+ },
+ {
+ X86::LOCK_AND8mi,
+ X86::LOCK_AND8mr,
+ X86::LOCK_AND16mi8,
+ X86::LOCK_AND16mi,
+ X86::LOCK_AND16mr,
+ X86::LOCK_AND32mi8,
+ X86::LOCK_AND32mi,
+ X86::LOCK_AND32mr,
+ X86::LOCK_AND64mi8,
+ X86::LOCK_AND64mi32,
+ X86::LOCK_AND64mr
+ },
+ {
+ X86::LOCK_XOR8mi,
+ X86::LOCK_XOR8mr,
+ X86::LOCK_XOR16mi8,
+ X86::LOCK_XOR16mi,
+ X86::LOCK_XOR16mr,
+ X86::LOCK_XOR32mi8,
+ X86::LOCK_XOR32mi,
+ X86::LOCK_XOR32mr,
+ X86::LOCK_XOR64mi8,
+ X86::LOCK_XOR64mi32,
+ X86::LOCK_XOR64mr
+ }
+};
+
+SDNode *X86DAGToDAGISel::SelectAtomicLoadArith(SDNode *Node, EVT NVT) {
+ if (Node->hasAnyUseOfValue(0))
+ return 0;
+
+ // Optimize common patterns for __sync_or_and_fetch and similar arith
+ // operations where the result is not used. This allows us to use the "lock"
+ // version of the arithmetic instruction.
+ // FIXME: Same as for 'add' and 'sub', try to merge those down here.
+ SDValue Chain = Node->getOperand(0);
+ SDValue Ptr = Node->getOperand(1);
+ SDValue Val = Node->getOperand(2);
+ SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
+ if (!SelectAddr(Node, Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
+ return 0;
+
+ // Which index into the table.
+ enum AtomicOpc Op;
+ switch (Node->getOpcode()) {
+ case ISD::ATOMIC_LOAD_OR:
+ Op = OR;
+ break;
+ case ISD::ATOMIC_LOAD_AND:
+ Op = AND;
+ break;
+ case ISD::ATOMIC_LOAD_XOR:
+ Op = XOR;
+ break;
+ default:
+ return 0;
+ }
+
+ bool isCN = false;
+ ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val);
+ if (CN && (int32_t)CN->getSExtValue() == CN->getSExtValue()) {
+ isCN = true;
+ Val = CurDAG->getTargetConstant(CN->getSExtValue(), NVT);
+ }
+
+ unsigned Opc = 0;
+ switch (NVT.getSimpleVT().SimpleTy) {
+ default: return 0;
+ case MVT::i8:
+ if (isCN)
+ Opc = AtomicOpcTbl[Op][ConstantI8];
+ else
+ Opc = AtomicOpcTbl[Op][I8];
+ break;
+ case MVT::i16:
+ if (isCN) {
+ if (immSext8(Val.getNode()))
+ Opc = AtomicOpcTbl[Op][SextConstantI16];
+ else
+ Opc = AtomicOpcTbl[Op][ConstantI16];
+ } else
+ Opc = AtomicOpcTbl[Op][I16];
+ break;
+ case MVT::i32:
+ if (isCN) {
+ if (immSext8(Val.getNode()))
+ Opc = AtomicOpcTbl[Op][SextConstantI32];
+ else
+ Opc = AtomicOpcTbl[Op][ConstantI32];
+ } else
+ Opc = AtomicOpcTbl[Op][I32];
+ break;
+ case MVT::i64:
+ Opc = AtomicOpcTbl[Op][I64];
+ if (isCN) {
+ if (immSext8(Val.getNode()))
+ Opc = AtomicOpcTbl[Op][SextConstantI64];
+ else if (i64immSExt32(Val.getNode()))
+ Opc = AtomicOpcTbl[Op][ConstantI64];
+ }
+ break;
+ }
+
+ assert(Opc != 0 && "Invalid arith lock transform!");
+
+ DebugLoc dl = Node->getDebugLoc();
+ SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ dl, NVT), 0);
+ MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+ MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
+ SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
+ SDValue Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7), 0);
+ cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
+ SDValue RetVals[] = { Undef, Ret };
+ return CurDAG->getMergeValues(RetVals, 2, 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) {
return RetVal;
break;
}
+ case ISD::ATOMIC_LOAD_XOR:
+ case ISD::ATOMIC_LOAD_AND:
+ case ISD::ATOMIC_LOAD_OR: {
+ SDNode *RetVal = SelectAtomicLoadArith(Node, NVT);
+ if (RetVal)
+ return RetVal;
+ break;
+ }
+ case ISD::AND:
+ case ISD::OR:
+ case ISD::XOR: {
+ // For operations of the form (x << C1) op C2, check if we can use a smaller
+ // encoding for C2 by transforming it into (x op (C2>>C1)) << C1.
+ SDValue N0 = Node->getOperand(0);
+ SDValue N1 = Node->getOperand(1);
+
+ if (N0->getOpcode() != ISD::SHL || !N0->hasOneUse())
+ break;
+
+ // i8 is unshrinkable, i16 should be promoted to i32.
+ if (NVT != MVT::i32 && NVT != MVT::i64)
+ break;
+
+ ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N1);
+ ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(N0->getOperand(1));
+ if (!Cst || !ShlCst)
+ break;
+
+ int64_t Val = Cst->getSExtValue();
+ uint64_t ShlVal = ShlCst->getZExtValue();
+
+ // Make sure that we don't change the operation by removing bits.
+ // This only matters for OR and XOR, AND is unaffected.
+ if (Opcode != ISD::AND && ((Val >> ShlVal) << ShlVal) != Val)
+ break;
+
+ unsigned ShlOp, Op = 0;
+ EVT CstVT = NVT;
+
+ // Check the minimum bitwidth for the new constant.
+ // TODO: AND32ri is the same as AND64ri32 with zext imm.
+ // TODO: MOV32ri+OR64r is cheaper than MOV64ri64+OR64rr
+ // TODO: Using 16 and 8 bit operations is also possible for or32 & xor32.
+ if (!isInt<8>(Val) && isInt<8>(Val >> ShlVal))
+ CstVT = MVT::i8;
+ else if (!isInt<32>(Val) && isInt<32>(Val >> ShlVal))
+ CstVT = MVT::i32;
+ // Bail if there is no smaller encoding.
+ if (NVT == CstVT)
+ break;
+
+ switch (NVT.getSimpleVT().SimpleTy) {
+ default: llvm_unreachable("Unsupported VT!");
+ case MVT::i32:
+ assert(CstVT == MVT::i8);
+ ShlOp = X86::SHL32ri;
+
+ switch (Opcode) {
+ case ISD::AND: Op = X86::AND32ri8; break;
+ case ISD::OR: Op = X86::OR32ri8; break;
+ case ISD::XOR: Op = X86::XOR32ri8; break;
+ }
+ break;
+ case MVT::i64:
+ assert(CstVT == MVT::i8 || CstVT == MVT::i32);
+ ShlOp = X86::SHL64ri;
+
+ switch (Opcode) {
+ case ISD::AND: Op = CstVT==MVT::i8? X86::AND64ri8 : X86::AND64ri32; break;
+ case ISD::OR: Op = CstVT==MVT::i8? X86::OR64ri8 : X86::OR64ri32; break;
+ case ISD::XOR: Op = CstVT==MVT::i8? X86::XOR64ri8 : X86::XOR64ri32; break;
+ }
+ break;
+ }
+
+ // Emit the smaller op and the shift.
+ SDValue NewCst = CurDAG->getTargetConstant(Val >> ShlVal, CstVT);
+ SDNode *New = CurDAG->getMachineNode(Op, dl, NVT, N0->getOperand(0),NewCst);
+ return CurDAG->SelectNodeTo(Node, ShlOp, NVT, SDValue(New, 0),
+ getI8Imm(ShlVal));
+ }
+ case X86ISD::UMUL: {
+ SDValue N0 = Node->getOperand(0);
+ SDValue N1 = Node->getOperand(1);
+
+ unsigned LoReg;
+ switch (NVT.getSimpleVT().SimpleTy) {
+ default: llvm_unreachable("Unsupported VT!");
+ case MVT::i8: LoReg = X86::AL; Opc = X86::MUL8r; break;
+ case MVT::i16: LoReg = X86::AX; Opc = X86::MUL16r; break;
+ case MVT::i32: LoReg = X86::EAX; Opc = X86::MUL32r; break;
+ case MVT::i64: LoReg = X86::RAX; Opc = X86::MUL64r; break;
+ }
+
+ SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
+ N0, SDValue()).getValue(1);
+
+ SDVTList VTs = CurDAG->getVTList(NVT, NVT, MVT::i32);
+ SDValue Ops[] = {N1, InFlag};
+ SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops, 2);
+
+ ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
+ ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
+ ReplaceUses(SDValue(Node, 2), SDValue(CNode, 2));
+ return NULL;
+ }
+
case ISD::SMUL_LOHI:
case ISD::UMUL_LOHI: {
SDValue N0 = Node->getOperand(0);
SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
InFlag };
SDNode *CNode =
- CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
+ CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops,
array_lengthof(Ops));
InFlag = SDValue(CNode, 1);
+
// Update the chain.
ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
} else {
- InFlag =
- SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
+ SDNode *CNode = CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag);
+ InFlag = SDValue(CNode, 0);
}
+ // Prevent use of AH in a REX instruction by referencing AX instead.
+ if (HiReg == X86::AH && Subtarget->is64Bit() &&
+ !SDValue(Node, 1).use_empty()) {
+ SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+ X86::AX, MVT::i16, InFlag);
+ InFlag = Result.getValue(2);
+ // Get the low part if needed. Don't use getCopyFromReg for aliasing
+ // registers.
+ if (!SDValue(Node, 0).use_empty())
+ ReplaceUses(SDValue(Node, 1),
+ CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
+
+ // Shift AX down 8 bits.
+ Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
+ Result,
+ CurDAG->getTargetConstant(8, MVT::i8)), 0);
+ // Then truncate it down to i8.
+ ReplaceUses(SDValue(Node, 1),
+ CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
+ }
// Copy the low half of the result, if it is needed.
if (!SDValue(Node, 0).use_empty()) {
SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
}
// Copy the high half of the result, if it is needed.
if (!SDValue(Node, 1).use_empty()) {
- SDValue Result;
- if (HiReg == X86::AH && Subtarget->is64Bit()) {
- // Prevent use of AH in a REX instruction by referencing AX instead.
- // Shift it down 8 bits.
- Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
- X86::AX, MVT::i16, InFlag);
- InFlag = Result.getValue(2);
- Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
- Result,
- CurDAG->getTargetConstant(8, MVT::i8)), 0);
- // Then truncate it down to i8.
- Result = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
- MVT::i8, Result);
- } else {
- Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
- HiReg, NVT, InFlag);
- InFlag = Result.getValue(2);
- }
+ SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+ HiReg, NVT, InFlag);
+ InFlag = Result.getValue(2);
ReplaceUses(SDValue(Node, 1), Result);
DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
}
-
+
return NULL;
}
if (TryFoldLoad(Node, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
Move =
- SDValue(CurDAG->getMachineNode(X86::MOVZX16rm8, dl, MVT::i16,
+ SDValue(CurDAG->getMachineNode(X86::MOVZX32rm8, dl, MVT::i32,
MVT::Other, Ops,
array_lengthof(Ops)), 0);
Chain = Move.getValue(1);
ReplaceUses(N0.getValue(1), Chain);
} else {
Move =
- SDValue(CurDAG->getMachineNode(X86::MOVZX16rr8, dl, MVT::i16, N0),0);
+ SDValue(CurDAG->getMachineNode(X86::MOVZX32rr8, dl, MVT::i32, N0),0);
Chain = CurDAG->getEntryNode();
}
- Chain = CurDAG->getCopyToReg(Chain, dl, X86::AX, Move, SDValue());
+ Chain = CurDAG->getCopyToReg(Chain, dl, X86::EAX, Move, SDValue());
InFlag = Chain.getValue(1);
} else {
InFlag =
if (isSigned && !signBitIsZero) {
// Sign extend the low part into the high part.
InFlag =
- SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Flag, InFlag),0);
+ SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Glue, InFlag),0);
} else {
// Zero out the high part, effectively zero extending the input.
SDValue ClrNode =
SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
InFlag };
SDNode *CNode =
- CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
+ CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops,
array_lengthof(Ops));
InFlag = SDValue(CNode, 1);
// Update the chain.
ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
} else {
InFlag =
- SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
+ SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag), 0);
}
+ // Prevent use of AH in a REX instruction by referencing AX instead.
+ // Shift it down 8 bits.
+ if (HiReg == X86::AH && Subtarget->is64Bit() &&
+ !SDValue(Node, 1).use_empty()) {
+ SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+ X86::AX, MVT::i16, InFlag);
+ InFlag = Result.getValue(2);
+
+ // If we also need AL (the quotient), get it by extracting a subreg from
+ // Result. The fast register allocator does not like multiple CopyFromReg
+ // nodes using aliasing registers.
+ if (!SDValue(Node, 0).use_empty())
+ ReplaceUses(SDValue(Node, 0),
+ CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
+
+ // Shift AX right by 8 bits instead of using AH.
+ Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
+ Result,
+ CurDAG->getTargetConstant(8, MVT::i8)),
+ 0);
+ ReplaceUses(SDValue(Node, 1),
+ CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
+ }
// Copy the division (low) result, if it is needed.
if (!SDValue(Node, 0).use_empty()) {
SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
}
// Copy the remainder (high) result, if it is needed.
if (!SDValue(Node, 1).use_empty()) {
- SDValue Result;
- if (HiReg == X86::AH && Subtarget->is64Bit()) {
- // Prevent use of AH in a REX instruction by referencing AX instead.
- // Shift it down 8 bits.
- Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
- X86::AX, MVT::i16, InFlag);
- InFlag = Result.getValue(2);
- Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
- Result,
- CurDAG->getTargetConstant(8, MVT::i8)),
- 0);
- // Then truncate it down to i8.
- Result = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
- MVT::i8, Result);
- } else {
- Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
- HiReg, NVT, InFlag);
- InFlag = Result.getValue(2);
- }
+ SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+ HiReg, NVT, InFlag);
+ InFlag = Result.getValue(2);
ReplaceUses(SDValue(Node, 1), Result);
DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
}
// Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
// use a smaller encoding.
- if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse())
+ if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
+ HasNoSignedComparisonUses(Node))
// Look past the truncate if CMP is the only use of it.
N0 = N0.getOperand(0);
- if (N0.getNode()->getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
+ if ((N0.getNode()->getOpcode() == ISD::AND ||
+ (N0.getResNo() == 0 && N0.getNode()->getOpcode() == X86ISD::AND)) &&
+ N0.getNode()->hasOneUse() &&
N0.getValueType() != MVT::i8 &&
X86::isZeroNode(N1)) {
ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getNode()->getOperand(1));
SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit_hi, dl,
MVT::i8, Reg);
- // Emit a testb. No special NOREX tricks are needed since there's
- // only one GPR operand!
- return CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
+ // Emit a testb. The EXTRACT_SUBREG becomes a COPY that can only
+ // target GR8_NOREX registers, so make sure the register class is
+ // forced.
+ return CurDAG->getMachineNode(X86::TEST8ri_NOREX, dl, MVT::i32,
Subreg, ShiftedImm);
}
}
break;
}
+ case ISD::STORE: {
+ // The DEC64m tablegen pattern is currently not able to match the case where
+ // the EFLAGS on the original DEC are used.
+ // we'll need to improve tablegen to allow flags to be transferred from a
+ // node in the pattern to the result node. probably with a new keyword
+ // for example, we have this
+ // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
+ // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
+ // (implicit EFLAGS)]>;
+ // but maybe need something like this
+ // def DEC64m : RI<0xFF, MRM1m, (outs), (ins i64mem:$dst), "dec{q}\t$dst",
+ // [(store (add (loadi64 addr:$dst), -1), addr:$dst),
+ // (transferrable EFLAGS)]>;
+ StoreSDNode *StoreNode = cast<StoreSDNode>(Node);
+ SDValue Chain = StoreNode->getOperand(0);
+ SDValue StoredVal = StoreNode->getOperand(1);
+ SDValue Address = StoreNode->getOperand(2);
+ SDValue Undef = StoreNode->getOperand(3);
+
+ if (StoreNode->getMemOperand()->getSize() != 8 ||
+ Undef->getOpcode() != ISD::UNDEF ||
+ Chain->getOpcode() != ISD::LOAD ||
+ StoredVal->getOpcode() != X86ISD::DEC ||
+ StoredVal.getResNo() != 0 ||
+ StoredVal->getOperand(0).getNode() != Chain.getNode())
+ break;
+
+ //OPC_CheckPredicate, 1, // Predicate_nontemporalstore
+ if (StoreNode->isNonTemporal())
+ break;
+
+ LoadSDNode *LoadNode = cast<LoadSDNode>(Chain.getNode());
+ if (LoadNode->getOperand(1) != Address ||
+ LoadNode->getOperand(2) != Undef)
+ break;
+
+ if (!ISD::isNormalLoad(LoadNode))
+ break;
+
+ if (!ISD::isNormalStore(StoreNode))
+ break;
+
+ // check load chain has only one use (from the store)
+ if (!Chain.hasOneUse())
+ break;
+
+ // Merge the input chains if they are not intra-pattern references.
+ SDValue InputChain = LoadNode->getOperand(0);
+
+ SDValue Base, Scale, Index, Disp, Segment;
+ if (!SelectAddr(LoadNode, LoadNode->getBasePtr(),
+ Base, Scale, Index, Disp, Segment))
+ break;
+
+ MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(2);
+ MemOp[0] = StoreNode->getMemOperand();
+ MemOp[1] = LoadNode->getMemOperand();
+ const SDValue Ops[] = { Base, Scale, Index, Disp, Segment, InputChain };
+ MachineSDNode *Result = CurDAG->getMachineNode(X86::DEC64m,
+ Node->getDebugLoc(),
+ MVT::i32, MVT::Other, Ops,
+ array_lengthof(Ops));
+ Result->setMemRefs(MemOp, MemOp + 2);
+
+ ReplaceUses(SDValue(StoreNode, 0), SDValue(Result, 1));
+ ReplaceUses(SDValue(StoredVal.getNode(), 1), SDValue(Result, 0));
+
+ return Result;
+ }
}
SDNode *ResNode = SelectCode(Node);
case 'v': // not offsetable ??
default: return true;
case 'm': // memory
- if (!SelectAddr(Op.getNode(), Op, Op0, Op1, Op2, Op3, Op4))
+ if (!SelectAddr(0, Op, Op0, Op1, Op2, Op3, Op4))
return true;
break;
}