//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "dagcombine"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
#include <algorithm>
using namespace llvm;
+#define DEBUG_TYPE "dagcombine"
+
STATISTIC(NodesCombined , "Number of dag nodes combined");
STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
namespace {
static cl::opt<bool>
CombinerAA("combiner-alias-analysis", cl::Hidden,
- cl::desc("Turn on alias analysis during testing"));
+ cl::desc("Enable DAG combiner alias-analysis heuristics"));
static cl::opt<bool>
CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
- cl::desc("Include global information in alias analysis"));
+ cl::desc("Enable DAG combiner's use of IR alias analysis"));
+
+ static cl::opt<bool>
+ UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true),
+ cl::desc("Enable DAG combiner's use of TBAA"));
+
+#ifndef NDEBUG
+ static cl::opt<std::string>
+ CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
+ cl::desc("Only use DAG-combiner alias analysis in this"
+ " function"));
+#endif
/// Hidden option to stress test load slicing, i.e., when this option
/// is enabled, load slicing bypasses most of its profitability guards.
/// now.
///
void AddUsersToWorkList(SDNode *N) {
- for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
- UI != UE; ++UI)
- AddToWorkList(*UI);
+ for (SDNode *Node : N->uses())
+ AddToWorkList(Node);
}
/// visit - call the node-specific routine that knows how to fold each
bool CombineToPostIndexedLoadStore(SDNode *N);
bool SliceUpLoad(SDNode *N);
+ /// \brief Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed
+ /// load.
+ ///
+ /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced.
+ /// \param InVecVT type of the input vector to EVE with bitcasts resolved.
+ /// \param EltNo index of the vector element to load.
+ /// \param OriginalLoad load that EVE came from to be replaced.
+ /// \returns EVE on success SDValue() on failure.
+ SDValue ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
+ SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad);
void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
SDValue visitSHL(SDNode *N);
SDValue visitSRA(SDNode *N);
SDValue visitSRL(SDNode *N);
+ SDValue visitRotate(SDNode *N);
SDValue visitCTLZ(SDNode *N);
SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
SDValue visitCTTZ(SDNode *N);
SDValue visitCONCAT_VECTORS(SDNode *N);
SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
SDValue visitVECTOR_SHUFFLE(SDNode *N);
+ SDValue visitINSERT_SUBVECTOR(SDNode *N);
SDValue XformToShuffleWithZero(SDNode *N);
SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
- SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
+ SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
bool NotExtCompare = false);
SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
SDLoc DL, bool foldBooleans = true);
+
+ bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
+ SDValue &CC) const;
+ bool isOneUseSetCC(SDValue N) const;
+
SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
unsigned HiOp);
SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
bool DemandHighBits = true);
SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
+ SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
+ SDValue InnerPos, SDValue InnerNeg,
+ unsigned PosOpcode, unsigned NegOpcode,
+ SDLoc DL);
SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
SDValue ReduceLoadWidth(SDNode *N);
SDValue ReduceLoadOpStoreWidth(SDNode *N);
/// isAlias - Return true if there is any possibility that the two addresses
/// overlap.
- bool isAlias(SDValue Ptr1, int64_t Size1, bool IsVolatile1,
- const Value *SrcValue1, int SrcValueOffset1,
- unsigned SrcValueAlign1,
- const MDNode *TBAAInfo1,
- SDValue Ptr2, int64_t Size2, bool IsVolatile2,
- const Value *SrcValue2, int SrcValueOffset2,
- unsigned SrcValueAlign2,
- const MDNode *TBAAInfo2) const;
-
- /// isAlias - Return true if there is any possibility that the two addresses
- /// overlap.
- bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1);
-
- /// FindAliasInfo - Extracts the relevant alias information from the memory
- /// node. Returns true if the operand was a load.
- bool FindAliasInfo(SDNode *N,
- SDValue &Ptr, int64_t &Size, bool &IsVolatile,
- const Value *&SrcValue, int &SrcValueOffset,
- unsigned &SrcValueAlignment,
- const MDNode *&TBAAInfo) const;
+ bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
/// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
/// looking for a better chain (aliasing node.)
/// \return True if some memory operations were changed.
bool MergeConsecutiveStores(StoreSDNode *N);
+ /// \brief Try to transform a truncation where C is a constant:
+ /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
+ ///
+ /// \p N needs to be a truncation and its first operand an AND. Other
+ /// requirements are checked by the function (e.g. that trunc is
+ /// single-use) and if missed an empty SDValue is returned.
+ SDValue distributeTruncateThroughAnd(SDNode *N);
+
public:
DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
: DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
explicit WorkListRemover(DAGCombiner &dc)
: SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
- virtual void NodeDeleted(SDNode *N, SDNode *E) {
+ void NodeDeleted(SDNode *N, SDNode *E) override {
DC.removeFromWorkList(N);
}
};
}
}
-
// isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
-// that selects between the values 1 and 0, making it equivalent to a setcc.
-// Also, set the incoming LHS, RHS, and CC references to the appropriate
-// nodes based on the type of node we are checking. This simplifies life a
-// bit for the callers.
-static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
- SDValue &CC) {
+// that selects between the target values used for true and false, making it
+// equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
+// the appropriate nodes based on the type of node we are checking. This
+// simplifies life a bit for the callers.
+bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
+ SDValue &CC) const {
if (N.getOpcode() == ISD::SETCC) {
LHS = N.getOperand(0);
RHS = N.getOperand(1);
CC = N.getOperand(2);
return true;
}
- if (N.getOpcode() == ISD::SELECT_CC &&
- N.getOperand(2).getOpcode() == ISD::Constant &&
- N.getOperand(3).getOpcode() == ISD::Constant &&
- cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 &&
- cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
- LHS = N.getOperand(0);
- RHS = N.getOperand(1);
- CC = N.getOperand(4);
- return true;
- }
- return false;
+
+ if (N.getOpcode() != ISD::SELECT_CC ||
+ !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
+ !TLI.isConstFalseVal(N.getOperand(3).getNode()))
+ return false;
+
+ LHS = N.getOperand(0);
+ RHS = N.getOperand(1);
+ CC = N.getOperand(4);
+ return true;
}
// isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
// one use. If this is true, it allows the users to invert the operation for
// free when it is profitable to do so.
-static bool isOneUseSetCC(SDValue N) {
+bool DAGCombiner::isOneUseSetCC(SDValue N) const {
SDValue N0, N1, N2;
if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
return true;
return false;
}
+/// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose
+/// elements are all the same constant or undefined.
+static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
+ BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
+ if (!C)
+ return false;
+
+ APInt SplatUndef;
+ unsigned SplatBitSize;
+ bool HasAnyUndefs;
+ EVT EltVT = N->getValueType(0).getVectorElementType();
+ return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
+ HasAnyUndefs) &&
+ EltVT.getSizeInBits() >= SplatBitSize);
+}
+
+// \brief Returns the SDNode if it is a constant BuildVector or constant.
+static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
+ if (isa<ConstantSDNode>(N))
+ return N.getNode();
+ BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
+ if(BV && BV->isConstant())
+ return BV;
+ return nullptr;
+}
+
+// \brief Returns the SDNode if it is a constant splat BuildVector or constant
+// int.
+static ConstantSDNode *isConstOrConstSplat(SDValue N) {
+ if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
+ return CN;
+
+ if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
+ ConstantSDNode *CN = BV->getConstantSplatValue();
+
+ // BuildVectors can truncate their operands. Ignore that case here.
+ if (CN && CN->getValueType(0) == N.getValueType().getScalarType())
+ return CN;
+ }
+
+ return nullptr;
+}
+
SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
SDValue N0, SDValue N1) {
EVT VT = N0.getValueType();
- if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
- if (isa<ConstantSDNode>(N1)) {
- // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
- SDValue OpNode =
- DAG.FoldConstantArithmetic(Opc, VT,
- cast<ConstantSDNode>(N0.getOperand(1)),
- cast<ConstantSDNode>(N1));
- return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
- }
- if (N0.hasOneUse()) {
- // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
- SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT,
- N0.getOperand(0), N1);
- AddToWorkList(OpNode.getNode());
- return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
- }
- }
-
- if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
- if (isa<ConstantSDNode>(N0)) {
- // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
- SDValue OpNode =
- DAG.FoldConstantArithmetic(Opc, VT,
- cast<ConstantSDNode>(N1.getOperand(1)),
- cast<ConstantSDNode>(N0));
- return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
- }
- if (N1.hasOneUse()) {
- // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
- SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT,
- N1.getOperand(0), N0);
- AddToWorkList(OpNode.getNode());
- return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
+ if (N0.getOpcode() == Opc) {
+ if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
+ if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
+ // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
+ SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R);
+ if (!OpNode.getNode())
+ return SDValue();
+ return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
+ }
+ if (N0.hasOneUse()) {
+ // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
+ // use
+ SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
+ if (!OpNode.getNode())
+ return SDValue();
+ AddToWorkList(OpNode.getNode());
+ return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
+ }
+ }
+ }
+
+ if (N1.getOpcode() == Opc) {
+ if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
+ if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
+ // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
+ SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L);
+ if (!OpNode.getNode())
+ return SDValue();
+ return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
+ }
+ if (N1.hasOneUse()) {
+ // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
+ // use
+ SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
+ if (!OpNode.getNode())
+ return SDValue();
+ AddToWorkList(OpNode.getNode());
+ return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
+ }
}
}
SDLoc dl(Op);
bool Replace = false;
SDValue NewOp = PromoteOperand(Op, PVT, Replace);
- if (NewOp.getNode() == 0)
+ if (!NewOp.getNode())
return SDValue();
AddToWorkList(NewOp.getNode());
SDLoc dl(Op);
bool Replace = false;
SDValue NewOp = PromoteOperand(Op, PVT, Replace);
- if (NewOp.getNode() == 0)
+ if (!NewOp.getNode())
return SDValue();
AddToWorkList(NewOp.getNode());
bool Replace0 = false;
SDValue N0 = Op.getOperand(0);
SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
- if (NN0.getNode() == 0)
+ if (!NN0.getNode())
return SDValue();
bool Replace1 = false;
NN1 = NN0;
else {
NN1 = PromoteOperand(N1, PVT, Replace1);
- if (NN1.getNode() == 0)
+ if (!NN1.getNode())
return SDValue();
}
N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
else
N0 = PromoteOperand(N0, PVT, Replace);
- if (N0.getNode() == 0)
+ if (!N0.getNode())
return SDValue();
AddToWorkList(N0.getNode());
SDValue RV = combine(N);
- if (RV.getNode() == 0)
+ if (!RV.getNode())
continue;
++NodesCombined;
case ISD::SHL: return visitSHL(N);
case ISD::SRA: return visitSRA(N);
case ISD::SRL: return visitSRL(N);
+ case ISD::ROTR:
+ case ISD::ROTL: return visitRotate(N);
case ISD::CTLZ: return visitCTLZ(N);
case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
case ISD::CTTZ: return visitCTTZ(N);
case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
+ case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
}
return SDValue();
}
SDValue RV = visit(N);
// If nothing happened, try a target-specific DAG combine.
- if (RV.getNode() == 0) {
+ if (!RV.getNode()) {
assert(N->getOpcode() != ISD::DELETED_NODE &&
"Node was deleted but visit returned NULL!");
}
// If nothing happened still, try promoting the operation.
- if (RV.getNode() == 0) {
+ if (!RV.getNode()) {
switch (N->getOpcode()) {
default: break;
case ISD::ADD:
// If N is a commutative binary node, try commuting it to enable more
// sdisel CSE.
- if (RV.getNode() == 0 &&
- SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
+ if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
N->getNumValues() == 1) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
// Constant operands are canonicalized to RHS.
if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
- SDValue Ops[] = { N1, N0 };
- SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
- Ops, 2);
+ SDValue Ops[] = {N1, N0};
+ SDNode *CSENode;
+ if (const BinaryWithFlagsSDNode *BinNode =
+ dyn_cast<BinaryWithFlagsSDNode>(N)) {
+ CSENode = DAG.getNodeIfExists(
+ N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
+ BinNode->hasNoSignedWrap(), BinNode->isExact());
+ } else {
+ CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
+ }
if (CSENode)
return SDValue(CSENode, 0);
}
Result = DAG.getEntryNode();
} else {
// New and improved token factor.
- Result = DAG.getNode(ISD::TokenFactor, SDLoc(N),
- MVT::Other, &Ops[0], Ops.size());
+ Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
}
// Don't add users to work list.
N0.getOperand(1));
// reassociate add
SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
- if (RADD.getNode() != 0)
+ if (RADD.getNode())
return RADD;
// fold ((0-A) + B) -> B-A
if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
if (VT.isInteger() && !VT.isVector()) {
APInt LHSZero, LHSOne;
APInt RHSZero, RHSOne;
- DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
+ DAG.computeKnownBits(N0, LHSZero, LHSOne);
if (LHSZero.getBoolValue()) {
- DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
+ DAG.computeKnownBits(N1, RHSZero, RHSOne);
// If all possibly-set bits on the LHS are clear on the RHS, return an OR.
// If all possibly-set bits on the RHS are clear on the LHS, return an OR.
- if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
- return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
+ if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
+ if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
+ return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
+ }
}
}
// fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
APInt LHSZero, LHSOne;
APInt RHSZero, RHSOne;
- DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
+ DAG.computeKnownBits(N0, LHSZero, LHSOne);
if (LHSZero.getBoolValue()) {
- DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
+ DAG.computeKnownBits(N1, RHSZero, RHSOne);
// If all possibly-set bits on the LHS are clear on the RHS, return an OR.
// If all possibly-set bits on the RHS are clear on the LHS, return an OR.
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
- ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 :
+ ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
EVT VT = N0.getValueType();
return SDValue();
}
-/// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose
-/// elements are all the same constant or undefined.
-static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
- BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
- if (!C)
- return false;
-
- APInt SplatUndef;
- unsigned SplatBitSize;
- bool HasAnyUndefs;
- EVT EltVT = N->getValueType(0).getVectorElementType();
- return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
- HasAnyUndefs) &&
- EltVT.getSizeInBits() >= SplatBitSize);
-}
-
SDValue DAGCombiner::visitMUL(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
} else {
- N0IsConst = dyn_cast<ConstantSDNode>(N0) != 0;
+ N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
: APInt();
- N1IsConst = dyn_cast<ConstantSDNode>(N1) != 0;
+ N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
: APInt();
}
// Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
// use.
{
- SDValue Sh(0,0), Y(0,0);
+ SDValue Sh(nullptr,0), Y(nullptr,0);
// Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
if (N0.getOpcode() == ISD::SHL &&
(isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
// reassociate mul
SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
- if (RMUL.getNode() != 0)
+ if (RMUL.getNode())
return RMUL;
return SDValue();
SDValue DAGCombiner::visitSDIV(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
- ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
- ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
+ ConstantSDNode *N0C = isConstOrConstSplat(N0);
+ ConstantSDNode *N1C = isConstOrConstSplat(N1);
EVT VT = N->getValueType(0);
// fold vector ops
return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
N0, N1);
}
+
// fold (sdiv X, pow2) -> simple ops after legalize
- if (N1C && !N1C->isNullValue() &&
- (N1C->getAPIntValue().isPowerOf2() ||
- (-N1C->getAPIntValue()).isPowerOf2())) {
+ if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
+ (-N1C->getAPIntValue()).isPowerOf2())) {
// If dividing by powers of two is cheap, then don't perform the following
// fold.
if (TLI.isPow2DivCheap())
unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
// Splat the sign bit into the register
- SDValue SGN = DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
- DAG.getConstant(VT.getSizeInBits()-1,
- getShiftAmountTy(N0.getValueType())));
+ SDValue SGN =
+ DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
+ DAG.getConstant(VT.getScalarSizeInBits() - 1,
+ getShiftAmountTy(N0.getValueType())));
AddToWorkList(SGN.getNode());
// Add (N0 < 0) ? abs2 - 1 : 0;
- SDValue SRL = DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
- DAG.getConstant(VT.getSizeInBits() - lg2,
- getShiftAmountTy(SGN.getValueType())));
+ SDValue SRL =
+ DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
+ DAG.getConstant(VT.getScalarSizeInBits() - lg2,
+ getShiftAmountTy(SGN.getValueType())));
SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
AddToWorkList(SRL.getNode());
AddToWorkList(ADD.getNode()); // Divide by pow2
return SRA;
AddToWorkList(SRA.getNode());
- return DAG.getNode(ISD::SUB, SDLoc(N), VT,
- DAG.getConstant(0, VT), SRA);
+ return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
}
// if integer divide is expensive and we satisfy the requirements, emit an
// alternate sequence.
- if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
+ if (N1C && !TLI.isIntDivCheap()) {
SDValue Op = BuildSDIV(N);
if (Op.getNode()) return Op;
}
SDValue DAGCombiner::visitUDIV(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
- ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
- ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
+ ConstantSDNode *N0C = isConstOrConstSplat(N0);
+ ConstantSDNode *N1C = isConstOrConstSplat(N1);
EVT VT = N->getValueType(0);
// fold vector ops
}
}
// fold (udiv x, c) -> alternate
- if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
+ if (N1C && !TLI.isIntDivCheap()) {
SDValue Op = BuildUDIV(N);
if (Op.getNode()) return Op;
}
SDValue DAGCombiner::visitSREM(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
- ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
- ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
+ ConstantSDNode *N0C = isConstOrConstSplat(N0);
+ ConstantSDNode *N1C = isConstOrConstSplat(N1);
EVT VT = N->getValueType(0);
// fold (srem c1, c2) -> c1%c2
SDValue DAGCombiner::visitUREM(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
- ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
- ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
+ ConstantSDNode *N0C = isConstOrConstSplat(N0);
+ ConstantSDNode *N1C = isConstOrConstSplat(N1);
EVT VT = N->getValueType(0);
// fold (urem c1, c2) -> c1%c2
bool HiExists = N->hasAnyUseOfValue(1);
if (!HiExists &&
(!LegalOperations ||
- TLI.isOperationLegal(LoOp, N->getValueType(0)))) {
+ TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
- N->op_begin(), N->getNumOperands());
+ ArrayRef<SDUse>(N->op_begin(), N->op_end()));
return CombineTo(N, Res, Res);
}
(!LegalOperations ||
TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
- N->op_begin(), N->getNumOperands());
+ ArrayRef<SDUse>(N->op_begin(), N->op_end()));
return CombineTo(N, Res, Res);
}
// If the two computed results can be simplified separately, separate them.
if (LoExists) {
SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
- N->op_begin(), N->getNumOperands());
+ ArrayRef<SDUse>(N->op_begin(), N->op_end()));
AddToWorkList(Lo.getNode());
SDValue LoOpt = combine(Lo.getNode());
if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
if (HiExists) {
SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
- N->op_begin(), N->getNumOperands());
+ ArrayRef<SDUse>(N->op_begin(), N->op_end()));
AddToWorkList(Hi.getNode());
SDValue HiOpt = combine(Hi.getNode());
if (HiOpt.getNode() && HiOpt != Hi &&
// The type-legalizer generates this pattern when loading illegal
// vector types from memory. In many cases this allows additional shuffle
// optimizations.
- if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
- N0.getOperand(1).getOpcode() == ISD::UNDEF &&
- N1.getOperand(1).getOpcode() == ISD::UNDEF) {
+ // There are other cases where moving the shuffle after the xor/and/or
+ // is profitable even if shuffles don't perform a swizzle.
+ // If both shuffles use the same mask, and both shuffles have the same first
+ // or second operand, then it might still be profitable to move the shuffle
+ // after the xor/and/or operation.
+ if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
- assert(N0.getOperand(0).getValueType() == N1.getOperand(1).getValueType() &&
+ assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
"Inputs to shuffles are not the same type");
- unsigned NumElts = VT.getVectorNumElements();
-
// Check that both shuffles use the same mask. The masks are known to be of
// the same length because the result vector type is the same.
- bool SameMask = true;
- for (unsigned i = 0; i != NumElts; ++i) {
- int Idx0 = SVN0->getMaskElt(i);
- int Idx1 = SVN1->getMaskElt(i);
- if (Idx0 != Idx1) {
- SameMask = false;
- break;
+ // Check also that shuffles have only one use to avoid introducing extra
+ // instructions.
+ if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
+ SVN0->getMask().equals(SVN1->getMask())) {
+ SDValue ShOp = N0->getOperand(1);
+
+ // Don't try to fold this node if it requires introducing a
+ // build vector of all zeros that might be illegal at this stage.
+ if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
+ if (!LegalTypes)
+ ShOp = DAG.getConstant(0, VT);
+ else
+ ShOp = SDValue();
}
- }
- if (SameMask) {
- SDValue Op = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
- N0.getOperand(0), N1.getOperand(0));
- AddToWorkList(Op.getNode());
- return DAG.getVectorShuffle(VT, SDLoc(N), Op,
- DAG.getUNDEF(VT), &SVN0->getMask()[0]);
+ // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
+ // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
+ // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
+ if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
+ SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
+ N0->getOperand(0), N1->getOperand(0));
+ AddToWorkList(NewNode.getNode());
+ return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
+ &SVN0->getMask()[0]);
+ }
+
+ // Don't try to fold this node if it requires introducing a
+ // build vector of all zeros that might be illegal at this stage.
+ ShOp = N0->getOperand(0);
+ if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
+ if (!LegalTypes)
+ ShOp = DAG.getConstant(0, VT);
+ else
+ ShOp = SDValue();
+ }
+
+ // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
+ // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
+ // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
+ if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
+ SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
+ N0->getOperand(1), N1->getOperand(1));
+ AddToWorkList(NewNode.getNode());
+ return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
+ &SVN0->getMask()[0]);
+ }
}
}
return DAG.getConstant(0, VT);
// reassociate and
SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
- if (RAND.getNode() != 0)
+ if (RAND.getNode())
return RAND;
// fold (and (or x, C), D) -> D if (C & D) == D
if (N1C && N0.getOpcode() == ISD::OR)
if (!TLI.isOperationLegal(ISD::BSWAP, VT))
return SDValue();
- SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
+ SmallVector<SDNode*,4> Parts(4, (SDNode*)nullptr);
// Look for either
// (or (or (and), (and)), (or (and), (and)))
// (or (or (or (and), (and)), (and)), (and))
return N0;
if (ISD::isBuildVectorAllOnes(N1.getNode()))
return N1;
+
+ // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
+ // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
+ // Do this only if the resulting shuffle is legal.
+ if (isa<ShuffleVectorSDNode>(N0) &&
+ isa<ShuffleVectorSDNode>(N1) &&
+ N0->getOperand(1) == N1->getOperand(1) &&
+ ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
+ bool CanFold = true;
+ unsigned NumElts = VT.getVectorNumElements();
+ const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
+ const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
+ // We construct two shuffle masks:
+ // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
+ // and N1 as the second operand.
+ // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
+ // and N0 as the second operand.
+ // We do this because OR is commutable and therefore there might be
+ // two ways to fold this node into a shuffle.
+ SmallVector<int,4> Mask1;
+ SmallVector<int,4> Mask2;
+
+ for (unsigned i = 0; i != NumElts && CanFold; ++i) {
+ int M0 = SV0->getMaskElt(i);
+ int M1 = SV1->getMaskElt(i);
+
+ // Both shuffle indexes are undef. Propagate Undef.
+ if (M0 < 0 && M1 < 0) {
+ Mask1.push_back(M0);
+ Mask2.push_back(M0);
+ continue;
+ }
+
+ if (M0 < 0 || M1 < 0 ||
+ (M0 < (int)NumElts && M1 < (int)NumElts) ||
+ (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
+ CanFold = false;
+ break;
+ }
+
+ Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
+ Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
+ }
+
+ if (CanFold) {
+ // Fold this sequence only if the resulting shuffle is 'legal'.
+ if (TLI.isShuffleMaskLegal(Mask1, VT))
+ return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
+ N1->getOperand(0), &Mask1[0]);
+ if (TLI.isShuffleMaskLegal(Mask2, VT))
+ return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
+ N0->getOperand(0), &Mask2[0]);
+ }
+ }
}
// fold (or x, undef) -> -1
// Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
SDValue BSwap = MatchBSwapHWord(N, N0, N1);
- if (BSwap.getNode() != 0)
+ if (BSwap.getNode())
return BSwap;
BSwap = MatchBSwapHWordLow(N, N0, N1);
- if (BSwap.getNode() != 0)
+ if (BSwap.getNode())
return BSwap;
// reassociate or
SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
- if (ROR.getNode() != 0)
+ if (ROR.getNode())
return ROR;
// Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
// iff (c1 & c2) == 0.
if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
isa<ConstantSDNode>(N0.getOperand(1))) {
ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
- if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
+ if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
+ SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1);
+ if (!COR.getNode())
+ return SDValue();
return DAG.getNode(ISD::AND, SDLoc(N), VT,
DAG.getNode(ISD::OR, SDLoc(N0), VT,
- N0.getOperand(0), N1),
- DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
+ N0.getOperand(0), N1), COR);
+ }
}
// fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
return false;
}
+// Return true if we can prove that, whenever Neg and Pos are both in the
+// range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
+// for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
+//
+// (or (shift1 X, Neg), (shift2 X, Pos))
+//
+// reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
+// in direction shift1 by Neg. The range [0, OpSize) means that we only need
+// to consider shift amounts with defined behavior.
+static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
+ // If OpSize is a power of 2 then:
+ //
+ // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
+ // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
+ //
+ // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
+ // for the stronger condition:
+ //
+ // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
+ //
+ // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
+ // we can just replace Neg with Neg' for the rest of the function.
+ //
+ // In other cases we check for the even stronger condition:
+ //
+ // Neg == OpSize - Pos [B]
+ //
+ // for all Neg and Pos. Note that the (or ...) then invokes undefined
+ // behavior if Pos == 0 (and consequently Neg == OpSize).
+ //
+ // We could actually use [A] whenever OpSize is a power of 2, but the
+ // only extra cases that it would match are those uninteresting ones
+ // where Neg and Pos are never in range at the same time. E.g. for
+ // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
+ // as well as (sub 32, Pos), but:
+ //
+ // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
+ //
+ // always invokes undefined behavior for 32-bit X.
+ //
+ // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
+ unsigned MaskLoBits = 0;
+ if (Neg.getOpcode() == ISD::AND &&
+ isPowerOf2_64(OpSize) &&
+ Neg.getOperand(1).getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
+ Neg = Neg.getOperand(0);
+ MaskLoBits = Log2_64(OpSize);
+ }
+
+ // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
+ if (Neg.getOpcode() != ISD::SUB)
+ return 0;
+ ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
+ if (!NegC)
+ return 0;
+ SDValue NegOp1 = Neg.getOperand(1);
+
+ // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
+ // Pos'. The truncation is redundant for the purpose of the equality.
+ if (MaskLoBits &&
+ Pos.getOpcode() == ISD::AND &&
+ Pos.getOperand(1).getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
+ Pos = Pos.getOperand(0);
+
+ // The condition we need is now:
+ //
+ // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
+ //
+ // If NegOp1 == Pos then we need:
+ //
+ // OpSize & Mask == NegC & Mask
+ //
+ // (because "x & Mask" is a truncation and distributes through subtraction).
+ APInt Width;
+ if (Pos == NegOp1)
+ Width = NegC->getAPIntValue();
+ // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
+ // Then the condition we want to prove becomes:
+ //
+ // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
+ //
+ // which, again because "x & Mask" is a truncation, becomes:
+ //
+ // NegC & Mask == (OpSize - PosC) & Mask
+ // OpSize & Mask == (NegC + PosC) & Mask
+ else if (Pos.getOpcode() == ISD::ADD &&
+ Pos.getOperand(0) == NegOp1 &&
+ Pos.getOperand(1).getOpcode() == ISD::Constant)
+ Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
+ NegC->getAPIntValue());
+ else
+ return false;
+
+ // Now we just need to check that OpSize & Mask == Width & Mask.
+ if (MaskLoBits)
+ // Opsize & Mask is 0 since Mask is Opsize - 1.
+ return Width.getLoBits(MaskLoBits) == 0;
+ return Width == OpSize;
+}
+
+// A subroutine of MatchRotate used once we have found an OR of two opposite
+// shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
+// to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
+// former being preferred if supported. InnerPos and InnerNeg are Pos and
+// Neg with outer conversions stripped away.
+SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
+ SDValue Neg, SDValue InnerPos,
+ SDValue InnerNeg, unsigned PosOpcode,
+ unsigned NegOpcode, SDLoc DL) {
+ // fold (or (shl x, (*ext y)),
+ // (srl x, (*ext (sub 32, y)))) ->
+ // (rotl x, y) or (rotr x, (sub 32, y))
+ //
+ // fold (or (shl x, (*ext (sub 32, y))),
+ // (srl x, (*ext y))) ->
+ // (rotr x, y) or (rotl x, (sub 32, y))
+ EVT VT = Shifted.getValueType();
+ if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
+ bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
+ return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
+ HasPos ? Pos : Neg).getNode();
+ }
+
+ return nullptr;
+}
+
// MatchRotate - Handle an 'or' of two operands. If this is one of the many
// idioms for rotate, and if the target supports rotation instructions, generate
// a rot[lr].
SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
// Must be a legal type. Expanded 'n promoted things won't work with rotates.
EVT VT = LHS.getValueType();
- if (!TLI.isTypeLegal(VT)) return 0;
+ if (!TLI.isTypeLegal(VT)) return nullptr;
// The target must have at least one rotate flavor.
bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
- if (!HasROTL && !HasROTR) return 0;
+ if (!HasROTL && !HasROTR) return nullptr;
// Match "(X shl/srl V1) & V2" where V2 may not be present.
SDValue LHSShift; // The shift.
SDValue LHSMask; // AND value if any.
if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
- return 0; // Not part of a rotate.
+ return nullptr; // Not part of a rotate.
SDValue RHSShift; // The shift.
SDValue RHSMask; // AND value if any.
if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
- return 0; // Not part of a rotate.
+ return nullptr; // Not part of a rotate.
if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
- return 0; // Not shifting the same value.
+ return nullptr; // Not shifting the same value.
if (LHSShift.getOpcode() == RHSShift.getOpcode())
- return 0; // Shifts must disagree.
+ return nullptr; // Shifts must disagree.
// Canonicalize shl to left side in a shl/srl pair.
if (RHSShift.getOpcode() == ISD::SHL) {
uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
if ((LShVal + RShVal) != OpSizeInBits)
- return 0;
+ return nullptr;
SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
// If there is a mask here, and we have a variable shift, we can't be sure
// that we're masking out the right stuff.
if (LHSMask.getNode() || RHSMask.getNode())
- return 0;
+ return nullptr;
// If the shift amount is sign/zext/any-extended just peel it off.
SDValue LExtOp0 = LHSShiftAmt;
RExtOp0 = RHSShiftAmt.getOperand(0);
}
- if (RExtOp0.getOpcode() == ISD::SUB && RExtOp0.getOperand(1) == LExtOp0) {
- // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
- // (rotl x, y)
- // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
- // (rotr x, (sub 32, y))
- if (ConstantSDNode *SUBC =
- dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) {
- if (SUBC->getAPIntValue() == OpSizeInBits) {
- return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT, LHSShiftArg,
- HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
- } else if (LHSShiftArg.getOpcode() == ISD::ZERO_EXTEND ||
- LHSShiftArg.getOpcode() == ISD::ANY_EXTEND) {
- // fold (or (shl (*ext x), (*ext y)),
- // (srl (*ext x), (*ext (sub 32, y)))) ->
- // (*ext (rotl x, y))
- // fold (or (shl (*ext x), (*ext y)),
- // (srl (*ext x), (*ext (sub 32, y)))) ->
- // (*ext (rotr x, (sub 32, y)))
- SDValue LArgExtOp0 = LHSShiftArg.getOperand(0);
- EVT LArgVT = LArgExtOp0.getValueType();
- bool HasROTRWithLArg = TLI.isOperationLegalOrCustom(ISD::ROTR, LArgVT);
- bool HasROTLWithLArg = TLI.isOperationLegalOrCustom(ISD::ROTL, LArgVT);
- if (HasROTRWithLArg || HasROTLWithLArg) {
- if (LArgVT.getSizeInBits() == SUBC->getAPIntValue()) {
- SDValue V =
- DAG.getNode(HasROTLWithLArg ? ISD::ROTL : ISD::ROTR, DL, LArgVT,
- LArgExtOp0, HasROTL ? LHSShiftAmt : RHSShiftAmt);
- return DAG.getNode(LHSShiftArg.getOpcode(), DL, VT, V).getNode();
- }
- }
- }
- }
- } else if (LExtOp0.getOpcode() == ISD::SUB &&
- RExtOp0 == LExtOp0.getOperand(1)) {
- // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
- // (rotr x, y)
- // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
- // (rotl x, (sub 32, y))
- if (ConstantSDNode *SUBC =
- dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) {
- if (SUBC->getAPIntValue() == OpSizeInBits) {
- return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT, LHSShiftArg,
- HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
- } else if (RHSShiftArg.getOpcode() == ISD::ZERO_EXTEND ||
- RHSShiftArg.getOpcode() == ISD::ANY_EXTEND) {
- // fold (or (shl (*ext x), (*ext (sub 32, y))),
- // (srl (*ext x), (*ext y))) ->
- // (*ext (rotl x, y))
- // fold (or (shl (*ext x), (*ext (sub 32, y))),
- // (srl (*ext x), (*ext y))) ->
- // (*ext (rotr x, (sub 32, y)))
- SDValue RArgExtOp0 = RHSShiftArg.getOperand(0);
- EVT RArgVT = RArgExtOp0.getValueType();
- bool HasROTRWithRArg = TLI.isOperationLegalOrCustom(ISD::ROTR, RArgVT);
- bool HasROTLWithRArg = TLI.isOperationLegalOrCustom(ISD::ROTL, RArgVT);
- if (HasROTRWithRArg || HasROTLWithRArg) {
- if (RArgVT.getSizeInBits() == SUBC->getAPIntValue()) {
- SDValue V =
- DAG.getNode(HasROTRWithRArg ? ISD::ROTR : ISD::ROTL, DL, RArgVT,
- RArgExtOp0, HasROTR ? RHSShiftAmt : LHSShiftAmt);
- return DAG.getNode(RHSShiftArg.getOpcode(), DL, VT, V).getNode();
- }
- }
- }
- }
- }
+ SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
+ LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
+ if (TryL)
+ return TryL;
+
+ SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
+ RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
+ if (TryR)
+ return TryR;
- return 0;
+ return nullptr;
}
SDValue DAGCombiner::visitXOR(SDNode *N) {
return N0;
// reassociate xor
SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
- if (RXOR.getNode() != 0)
+ if (RXOR.getNode())
return RXOR;
// fold !(x cc y) -> (x !cc y)
/// visitShiftByConstant - Handle transforms common to the three shifts, when
/// the shift amount is a constant.
-SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
+SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
+ // We can't and shouldn't fold opaque constants.
+ if (Amt->isOpaque())
+ return SDValue();
+
SDNode *LHS = N->getOperand(0).getNode();
if (!LHS->hasOneUse()) return SDValue();
break;
}
- // We require the RHS of the binop to be a constant as well.
+ // We require the RHS of the binop to be a constant and not opaque as well.
ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
- if (!BinOpCst) return SDValue();
+ if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
// FIXME: disable this unless the input to the binop is a shift by a constant.
// If it is not a shift, it pessimizes some common cases like:
return SDValue();
}
+ if (!TLI.isDesirableToCommuteWithShift(LHS))
+ return SDValue();
+
// Fold the constants, shifting the binop RHS by the shift amount.
SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
N->getValueType(0),
LHS->getOperand(1), N->getOperand(1));
+ assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
// Create the new shift.
SDValue NewShift = DAG.getNode(N->getOpcode(),
return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
}
+SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
+ assert(N->getOpcode() == ISD::TRUNCATE);
+ assert(N->getOperand(0).getOpcode() == ISD::AND);
+
+ // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
+ if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
+ SDValue N01 = N->getOperand(0).getOperand(1);
+
+ if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
+ EVT TruncVT = N->getValueType(0);
+ SDValue N00 = N->getOperand(0).getOperand(0);
+ APInt TruncC = N01C->getAPIntValue();
+ TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
+
+ return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
+ DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
+ DAG.getConstant(TruncC, TruncVT));
+ }
+ }
+
+ return SDValue();
+}
+
+SDValue DAGCombiner::visitRotate(SDNode *N) {
+ // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
+ if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
+ N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
+ SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
+ if (NewOp1.getNode())
+ return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
+ N->getOperand(0), NewOp1);
+ }
+ return SDValue();
+}
+
SDValue DAGCombiner::visitSHL(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
EVT VT = N0.getValueType();
- unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
+ unsigned OpSizeInBits = VT.getScalarSizeInBits();
// fold vector ops
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
+
+ BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
+ // If setcc produces all-one true value then:
+ // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
+ if (N1CV && N1CV->isConstant()) {
+ if (N0.getOpcode() == ISD::AND &&
+ TLI.getBooleanContents(true) ==
+ TargetLowering::ZeroOrNegativeOneBooleanContent) {
+ SDValue N00 = N0->getOperand(0);
+ SDValue N01 = N0->getOperand(1);
+ BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
+
+ if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC) {
+ SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV);
+ if (C.getNode())
+ return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
+ }
+ } else {
+ N1C = isConstOrConstSplat(N1);
+ }
+ }
}
// fold (shl c1, c2) -> c1<<c2
return DAG.getConstant(0, VT);
// fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
if (N1.getOpcode() == ISD::TRUNCATE &&
- N1.getOperand(0).getOpcode() == ISD::AND &&
- N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
- SDValue N101 = N1.getOperand(0).getOperand(1);
- if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
- EVT TruncVT = N1.getValueType();
- SDValue N100 = N1.getOperand(0).getOperand(0);
- APInt TruncC = N101C->getAPIntValue();
- TruncC = TruncC.trunc(TruncVT.getSizeInBits());
- return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
- DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
- DAG.getNode(ISD::TRUNCATE,
- SDLoc(N),
- TruncVT, N100),
- DAG.getConstant(TruncC, TruncVT)));
- }
+ N1.getOperand(0).getOpcode() == ISD::AND) {
+ SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
+ if (NewOp1.getNode())
+ return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
}
if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
return SDValue(N, 0);
// fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
- if (N1C && N0.getOpcode() == ISD::SHL &&
- N0.getOperand(1).getOpcode() == ISD::Constant) {
- uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
- uint64_t c2 = N1C->getZExtValue();
- if (c1 + c2 >= OpSizeInBits)
- return DAG.getConstant(0, VT);
- return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
- DAG.getConstant(c1 + c2, N1.getValueType()));
+ if (N1C && N0.getOpcode() == ISD::SHL) {
+ if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
+ uint64_t c1 = N0C1->getZExtValue();
+ uint64_t c2 = N1C->getZExtValue();
+ if (c1 + c2 >= OpSizeInBits)
+ return DAG.getConstant(0, VT);
+ return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
+ DAG.getConstant(c1 + c2, N1.getValueType()));
+ }
}
// fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
N0.getOpcode() == ISD::ANY_EXTEND ||
N0.getOpcode() == ISD::SIGN_EXTEND) &&
- N0.getOperand(0).getOpcode() == ISD::SHL &&
- isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
- uint64_t c1 =
- cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
- uint64_t c2 = N1C->getZExtValue();
- EVT InnerShiftVT = N0.getOperand(0).getValueType();
- uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
- if (c2 >= OpSizeInBits - InnerShiftSize) {
- if (c1 + c2 >= OpSizeInBits)
- return DAG.getConstant(0, VT);
- return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
- DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
- N0.getOperand(0)->getOperand(0)),
- DAG.getConstant(c1 + c2, N1.getValueType()));
+ N0.getOperand(0).getOpcode() == ISD::SHL) {
+ SDValue N0Op0 = N0.getOperand(0);
+ if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
+ uint64_t c1 = N0Op0C1->getZExtValue();
+ uint64_t c2 = N1C->getZExtValue();
+ EVT InnerShiftVT = N0Op0.getValueType();
+ uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
+ if (c2 >= OpSizeInBits - InnerShiftSize) {
+ if (c1 + c2 >= OpSizeInBits)
+ return DAG.getConstant(0, VT);
+ return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
+ DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
+ N0Op0->getOperand(0)),
+ DAG.getConstant(c1 + c2, N1.getValueType()));
+ }
}
}
// Only fold this if the inner zext has no other uses to avoid increasing
// the total number of instructions.
if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
- N0.getOperand(0).getOpcode() == ISD::SRL &&
- isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
- uint64_t c1 =
- cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
- if (c1 < VT.getSizeInBits()) {
- uint64_t c2 = N1C->getZExtValue();
- if (c1 == c2) {
- SDValue NewOp0 = N0.getOperand(0);
- EVT CountVT = NewOp0.getOperand(1).getValueType();
- SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
- NewOp0, DAG.getConstant(c2, CountVT));
- AddToWorkList(NewSHL.getNode());
- return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
+ N0.getOperand(0).getOpcode() == ISD::SRL) {
+ SDValue N0Op0 = N0.getOperand(0);
+ if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
+ uint64_t c1 = N0Op0C1->getZExtValue();
+ if (c1 < VT.getScalarSizeInBits()) {
+ uint64_t c2 = N1C->getZExtValue();
+ if (c1 == c2) {
+ SDValue NewOp0 = N0.getOperand(0);
+ EVT CountVT = NewOp0.getOperand(1).getValueType();
+ SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
+ NewOp0, DAG.getConstant(c2, CountVT));
+ AddToWorkList(NewSHL.getNode());
+ return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
+ }
}
}
}
// (and (srl x, (sub c1, c2), MASK)
// Only fold this if the inner shift has no other uses -- if it does, folding
// this will increase the total number of instructions.
- if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
- N0.getOperand(1).getOpcode() == ISD::Constant) {
- uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
- if (c1 < VT.getSizeInBits()) {
- uint64_t c2 = N1C->getZExtValue();
- APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
- VT.getSizeInBits() - c1);
- SDValue Shift;
- if (c2 > c1) {
- Mask = Mask.shl(c2-c1);
- Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
- DAG.getConstant(c2-c1, N1.getValueType()));
- } else {
- Mask = Mask.lshr(c1-c2);
- Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
- DAG.getConstant(c1-c2, N1.getValueType()));
+ if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
+ if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
+ uint64_t c1 = N0C1->getZExtValue();
+ if (c1 < OpSizeInBits) {
+ uint64_t c2 = N1C->getZExtValue();
+ APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
+ SDValue Shift;
+ if (c2 > c1) {
+ Mask = Mask.shl(c2 - c1);
+ Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
+ DAG.getConstant(c2 - c1, N1.getValueType()));
+ } else {
+ Mask = Mask.lshr(c1 - c2);
+ Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
+ DAG.getConstant(c1 - c2, N1.getValueType()));
+ }
+ return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
+ DAG.getConstant(Mask, VT));
}
- return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
- DAG.getConstant(Mask, VT));
}
}
// fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
+ unsigned BitSize = VT.getScalarSizeInBits();
SDValue HiBitsMask =
- DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
- VT.getSizeInBits() -
- N1C->getZExtValue()),
- VT);
+ DAG.getConstant(APInt::getHighBitsSet(BitSize,
+ BitSize - N1C->getZExtValue()), VT);
return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
HiBitsMask);
}
if (N1C) {
- SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
+ SDValue NewSHL = visitShiftByConstant(N, N1C);
if (NewSHL.getNode())
return NewSHL;
}
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
+
+ N1C = isConstOrConstSplat(N1);
}
// fold (sra c1, c2) -> (sra c1, c2)
// fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
if (N1C && N0.getOpcode() == ISD::SRA) {
- if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
+ if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
- if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
+ if (Sum >= OpSizeInBits)
+ Sum = OpSizeInBits - 1;
return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
- DAG.getConstant(Sum, N1C->getValueType(0)));
+ DAG.getConstant(Sum, N1.getValueType()));
}
}
// result_size - n != m.
// If truncate is free for the target sext(shl) is likely to result in better
// code.
- if (N0.getOpcode() == ISD::SHL) {
+ if (N0.getOpcode() == ISD::SHL && N1C) {
// Get the two constanst of the shifts, CN0 = m, CN = n.
- const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
- if (N01C && N1C) {
+ const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
+ if (N01C) {
+ LLVMContext &Ctx = *DAG.getContext();
// Determine what the truncate's result bitsize and type would be.
- EVT TruncVT =
- EVT::getIntegerVT(*DAG.getContext(),
- OpSizeInBits - N1C->getZExtValue());
+ EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
+
+ if (VT.isVector())
+ TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
+
// Determine the residual right-shift amount.
signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
// fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
if (N1.getOpcode() == ISD::TRUNCATE &&
- N1.getOperand(0).getOpcode() == ISD::AND &&
- N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
- SDValue N101 = N1.getOperand(0).getOperand(1);
- if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
- EVT TruncVT = N1.getValueType();
- SDValue N100 = N1.getOperand(0).getOperand(0);
- APInt TruncC = N101C->getAPIntValue();
- TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
- return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
- DAG.getNode(ISD::AND, SDLoc(N),
- TruncVT,
- DAG.getNode(ISD::TRUNCATE,
- SDLoc(N),
- TruncVT, N100),
- DAG.getConstant(TruncC, TruncVT)));
- }
- }
-
- // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
+ N1.getOperand(0).getOpcode() == ISD::AND) {
+ SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
+ if (NewOp1.getNode())
+ return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
+ }
+
+ // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
// if c1 is equal to the number of bits the trunc removes
if (N0.getOpcode() == ISD::TRUNCATE &&
(N0.getOperand(0).getOpcode() == ISD::SRL ||
N0.getOperand(0).getOpcode() == ISD::SRA) &&
N0.getOperand(0).hasOneUse() &&
N0.getOperand(0).getOperand(1).hasOneUse() &&
- N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
- EVT LargeVT = N0.getOperand(0).getValueType();
- ConstantSDNode *LargeShiftAmt =
- cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
-
- if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
- LargeShiftAmt->getZExtValue()) {
- SDValue Amt =
- DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
- getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
- SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
- N0.getOperand(0).getOperand(0), Amt);
- return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
+ N1C) {
+ SDValue N0Op0 = N0.getOperand(0);
+ if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
+ unsigned LargeShiftVal = LargeShift->getZExtValue();
+ EVT LargeVT = N0Op0.getValueType();
+
+ if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
+ SDValue Amt =
+ DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
+ getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
+ SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
+ N0Op0.getOperand(0), Amt);
+ return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
+ }
}
}
return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
if (N1C) {
- SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
+ SDValue NewSRA = visitShiftByConstant(N, N1C);
if (NewSRA.getNode())
return NewSRA;
}
if (VT.isVector()) {
SDValue FoldedVOp = SimplifyVBinOp(N);
if (FoldedVOp.getNode()) return FoldedVOp;
+
+ N1C = isConstOrConstSplat(N1);
}
// fold (srl c1, c2) -> c1 >>u c2
return DAG.getConstant(0, VT);
// fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
- if (N1C && N0.getOpcode() == ISD::SRL &&
- N0.getOperand(1).getOpcode() == ISD::Constant) {
- uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
- uint64_t c2 = N1C->getZExtValue();
- if (c1 + c2 >= OpSizeInBits)
- return DAG.getConstant(0, VT);
- return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
- DAG.getConstant(c1 + c2, N1.getValueType()));
+ if (N1C && N0.getOpcode() == ISD::SRL) {
+ if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
+ uint64_t c1 = N01C->getZExtValue();
+ uint64_t c2 = N1C->getZExtValue();
+ if (c1 + c2 >= OpSizeInBits)
+ return DAG.getConstant(0, VT);
+ return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
+ DAG.getConstant(c1 + c2, N1.getValueType()));
+ }
}
// fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
}
// fold (srl (shl x, c), c) -> (and x, cst2)
- if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
- N0.getValueSizeInBits() <= 64) {
- uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
- return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
- DAG.getConstant(~0ULL >> ShAmt, VT));
+ if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
+ unsigned BitSize = N0.getScalarValueSizeInBits();
+ if (BitSize <= 64) {
+ uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
+ return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
+ DAG.getConstant(~0ULL >> ShAmt, VT));
+ }
}
// fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
// Shifting in all undef bits?
EVT SmallVT = N0.getOperand(0).getValueType();
- if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
+ unsigned BitSize = SmallVT.getScalarSizeInBits();
+ if (N1C->getZExtValue() >= BitSize)
return DAG.getUNDEF(VT);
if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
N0.getOperand(0),
DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
AddToWorkList(SmallShift.getNode());
- APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits()).lshr(ShiftAmt);
+ APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
return DAG.getNode(ISD::AND, SDLoc(N), VT,
DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
DAG.getConstant(Mask, VT));
// fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
// bit, which is unmodified by sra.
- if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
+ if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
if (N0.getOpcode() == ISD::SRA)
return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
}
// fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
if (N1C && N0.getOpcode() == ISD::CTLZ &&
- N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
+ N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
APInt KnownZero, KnownOne;
- DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
+ DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
// If any of the input bits are KnownOne, then the input couldn't be all
// zeros, thus the result of the srl will always be zero.
// fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
if (N1.getOpcode() == ISD::TRUNCATE &&
- N1.getOperand(0).getOpcode() == ISD::AND &&
- N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
- SDValue N101 = N1.getOperand(0).getOperand(1);
- if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
- EVT TruncVT = N1.getValueType();
- SDValue N100 = N1.getOperand(0).getOperand(0);
- APInt TruncC = N101C->getAPIntValue();
- TruncC = TruncC.trunc(TruncVT.getSizeInBits());
- return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
- DAG.getNode(ISD::AND, SDLoc(N),
- TruncVT,
- DAG.getNode(ISD::TRUNCATE,
- SDLoc(N),
- TruncVT, N100),
- DAG.getConstant(TruncC, TruncVT)));
- }
+ N1.getOperand(0).getOpcode() == ISD::AND) {
+ SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
+ if (NewOp1.getNode())
+ return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
}
// fold operands of srl based on knowledge that the low bits are not
return SDValue(N, 0);
if (N1C) {
- SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
+ SDValue NewSRL = visitShiftByConstant(N, N1C);
if (NewSRL.getNode())
return NewSRL;
}
std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
SDLoc DL(N);
EVT LoVT, HiVT;
- llvm::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
// Split the inputs.
SDValue Lo, Hi, LL, LH, RL, RH;
- llvm::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
- llvm::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
+ std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
+ std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
return std::make_pair(Lo, Hi);
}
+// This function assumes all the vselect's arguments are CONCAT_VECTOR
+// nodes and that the condition is a BV of ConstantSDNodes (or undefs).
+static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
+ SDLoc dl(N);
+ SDValue Cond = N->getOperand(0);
+ SDValue LHS = N->getOperand(1);
+ SDValue RHS = N->getOperand(2);
+ MVT VT = N->getSimpleValueType(0);
+ int NumElems = VT.getVectorNumElements();
+ assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
+ RHS.getOpcode() == ISD::CONCAT_VECTORS &&
+ Cond.getOpcode() == ISD::BUILD_VECTOR);
+
+ // We're sure we have an even number of elements due to the
+ // concat_vectors we have as arguments to vselect.
+ // Skip BV elements until we find one that's not an UNDEF
+ // After we find an UNDEF element, keep looping until we get to half the
+ // length of the BV and see if all the non-undef nodes are the same.
+ ConstantSDNode *BottomHalf = nullptr;
+ for (int i = 0; i < NumElems / 2; ++i) {
+ if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
+ continue;
+
+ if (BottomHalf == nullptr)
+ BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
+ else if (Cond->getOperand(i).getNode() != BottomHalf)
+ return SDValue();
+ }
+
+ // Do the same for the second half of the BuildVector
+ ConstantSDNode *TopHalf = nullptr;
+ for (int i = NumElems / 2; i < NumElems; ++i) {
+ if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
+ continue;
+
+ if (TopHalf == nullptr)
+ TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
+ else if (Cond->getOperand(i).getNode() != TopHalf)
+ return SDValue();
+ }
+
+ assert(TopHalf && BottomHalf &&
+ "One half of the selector was all UNDEFs and the other was all the "
+ "same value. This should have been addressed before this function.");
+ return DAG.getNode(
+ ISD::CONCAT_VECTORS, dl, VT,
+ BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
+ TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
+}
+
SDValue DAGCombiner::visitVSELECT(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
return SDValue();
SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
- llvm::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
- llvm::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
- llvm::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
+ std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
+ std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
+ std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
}
+ // Fold (vselect (build_vector all_ones), N1, N2) -> N1
+ if (ISD::isBuildVectorAllOnes(N0.getNode()))
+ return N1;
+ // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
+ if (ISD::isBuildVectorAllZeros(N0.getNode()))
+ return N2;
+
+ // The ConvertSelectToConcatVector function is assuming both the above
+ // checks for (vselect (build_vector all{ones,zeros) ...) have been made
+ // and addressed.
+ if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
+ N2.getOpcode() == ISD::CONCAT_VECTORS &&
+ ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
+ SDValue CV = ConvertSelectToConcatVector(N, DAG);
+ if (CV.getNode())
+ return CV;
+ }
+
return SDValue();
}
SDLoc(N));
}
+// tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
+// dag node into a ConstantSDNode or a build_vector of constants.
+// This function is called by the DAGCombiner when visiting sext/zext/aext
+// dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
+// Vector extends are not folded if operations are legal; this is to
+// avoid introducing illegal build_vector dag nodes.
+static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
+ SelectionDAG &DAG, bool LegalTypes,
+ bool LegalOperations) {
+ unsigned Opcode = N->getOpcode();
+ SDValue N0 = N->getOperand(0);
+ EVT VT = N->getValueType(0);
+
+ assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
+ Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
+
+ // fold (sext c1) -> c1
+ // fold (zext c1) -> c1
+ // fold (aext c1) -> c1
+ if (isa<ConstantSDNode>(N0))
+ return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
+
+ // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
+ // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
+ // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
+ EVT SVT = VT.getScalarType();
+ if (!(VT.isVector() &&
+ (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
+ ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
+ return nullptr;
+
+ // We can fold this node into a build_vector.
+ unsigned VTBits = SVT.getSizeInBits();
+ unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
+ unsigned ShAmt = VTBits - EVTBits;
+ SmallVector<SDValue, 8> Elts;
+ unsigned NumElts = N0->getNumOperands();
+ SDLoc DL(N);
+
+ for (unsigned i=0; i != NumElts; ++i) {
+ SDValue Op = N0->getOperand(i);
+ if (Op->getOpcode() == ISD::UNDEF) {
+ Elts.push_back(DAG.getUNDEF(SVT));
+ continue;
+ }
+
+ ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
+ const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
+ if (Opcode == ISD::SIGN_EXTEND)
+ Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
+ SVT));
+ else
+ Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
+ SVT));
+ }
+
+ return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
+}
+
// ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
// "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
// transformation. Returns true if extension are possible and the above
}
Ops.push_back(SetCC->getOperand(2));
- CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
- &Ops[0], Ops.size()));
+ CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
}
}
SDValue N0 = N->getOperand(0);
EVT VT = N->getValueType(0);
- // fold (sext c1) -> c1
- if (isa<ConstantSDNode>(N0))
- return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N0);
+ if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
+ LegalOperations))
+ return SDValue(Res, 0);
// fold (sext (sext x)) -> (sext x)
// fold (sext (aext x)) -> (sext x)
// on vectors in one instruction. We only perform this transformation on
// scalars.
if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
+ ISD::isUNINDEXEDLoad(N0.getNode()) &&
((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
bool DoXform = true;
TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
(!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
- if (LN0->getExtensionType() != ISD::ZEXTLOAD) {
+ if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
bool DoXform = true;
SmallVector<SDNode*, 4> SetCCs;
if (!N0.hasOneUse())
}
}
- // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
+ // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
unsigned ElementWidth = VT.getScalarType().getSizeInBits();
SDValue NegOne =
DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
NegOne, DAG.getConstant(0, VT),
cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
if (SCC.getNode()) return SCC;
- if (!VT.isVector() &&
- (!LegalOperations ||
- TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(VT)))) {
- return DAG.getSelect(SDLoc(N), VT,
- DAG.getSetCC(SDLoc(N),
- getSetCCResultType(VT),
- N0.getOperand(0), N0.getOperand(1),
- cast<CondCodeSDNode>(N0.getOperand(2))->get()),
- NegOne, DAG.getConstant(0, VT));
+
+ if (!VT.isVector()) {
+ EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
+ if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
+ SDLoc DL(N);
+ ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
+ SDValue SetCC = DAG.getSetCC(DL,
+ SetCCVT,
+ N0.getOperand(0), N0.getOperand(1), CC);
+ EVT SelectVT = getSetCCResultType(VT);
+ return DAG.getSelect(DL, VT,
+ DAG.getSExtOrTrunc(SetCC, DL, SelectVT),
+ NegOne, DAG.getConstant(0, VT));
+
+ }
}
}
// isTruncateOf - If N is a truncate of some other value, return true, record
// the value being truncated in Op and which of Op's bits are zero in KnownZero.
// This function computes KnownZero to avoid a duplicated call to
-// ComputeMaskedBits in the caller.
+// computeKnownBits in the caller.
static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
APInt &KnownZero) {
APInt KnownOne;
if (N->getOpcode() == ISD::TRUNCATE) {
Op = N->getOperand(0);
- DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
+ DAG.computeKnownBits(Op, KnownZero, KnownOne);
return true;
}
else
return false;
- DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
+ DAG.computeKnownBits(Op, KnownZero, KnownOne);
if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
return false;
SDValue N0 = N->getOperand(0);
EVT VT = N->getValueType(0);
- // fold (zext c1) -> c1
- if (isa<ConstantSDNode>(N0))
- return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
+ if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
+ LegalOperations))
+ return SDValue(Res, 0);
+
// fold (zext (zext x)) -> (zext x)
// fold (zext (aext x)) -> (zext x)
if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
// on vectors in one instruction. We only perform this transformation on
// scalars.
if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
+ ISD::isUNINDEXEDLoad(N0.getNode()) &&
((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
bool DoXform = true;
TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
(!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
- if (LN0->getExtensionType() != ISD::SEXTLOAD) {
+ if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
bool DoXform = true;
SmallVector<SDNode*, 4> SetCCs;
if (!N0.hasOneUse())
}
if (N0.getOpcode() == ISD::SETCC) {
- if (!LegalOperations && VT.isVector()) {
+ if (!LegalOperations && VT.isVector() &&
+ N0.getValueType().getVectorElementType() == MVT::i1) {
+ EVT N0VT = N0.getOperand(0).getValueType();
+ if (getSetCCResultType(N0VT) == N0.getValueType())
+ return SDValue();
+
// zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
// Only do this before legalize for now.
- EVT N0VT = N0.getOperand(0).getValueType();
EVT EltVT = VT.getVectorElementType();
SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
DAG.getConstant(1, EltVT));
N0.getOperand(1),
cast<CondCodeSDNode>(N0.getOperand(2))->get()),
DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
- &OneOps[0], OneOps.size()));
+ OneOps));
// If the desired elements are smaller or larger than the source
// elements we can use a matching integer vector type and then
cast<CondCodeSDNode>(N0.getOperand(2))->get());
return DAG.getNode(ISD::AND, SDLoc(N), VT,
DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
- DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
- &OneOps[0], OneOps.size()));
+ DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
}
// zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
SDValue N0 = N->getOperand(0);
EVT VT = N->getValueType(0);
- // fold (aext c1) -> c1
- if (isa<ConstantSDNode>(N0))
- return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, N0);
+ if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
+ LegalOperations))
+ return SDValue(Res, 0);
+
// fold (aext (aext x)) -> (aext x)
// fold (aext (zext x)) -> (zext x)
// fold (aext (sext x)) -> (sext x)
// on vectors in one instruction. We only perform this transformation on
// scalars.
if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
+ ISD::isUNINDEXEDLoad(N0.getNode()) &&
((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
bool DoXform = true;
!ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
N0.hasOneUse()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
+ ISD::LoadExtType ExtType = LN0->getExtensionType();
EVT MemVT = LN0->getMemoryVT();
- SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(N),
- VT, LN0->getChain(), LN0->getBasePtr(),
- MemVT, LN0->getMemOperand());
- CombineTo(N, ExtLoad);
- CombineTo(N0.getNode(),
- DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
- N0.getValueType(), ExtLoad),
- ExtLoad.getValue(1));
- return SDValue(N, 0); // Return N so it doesn't get rechecked!
+ if (!LegalOperations || TLI.isLoadExtLegal(ExtType, MemVT)) {
+ SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
+ VT, LN0->getChain(), LN0->getBasePtr(),
+ MemVT, LN0->getMemOperand());
+ CombineTo(N, ExtLoad);
+ CombineTo(N0.getNode(),
+ DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
+ N0.getValueType(), ExtLoad),
+ ExtLoad.getValue(1));
+ return SDValue(N, 0); // Return N so it doesn't get rechecked!
+ }
}
if (N0.getOpcode() == ISD::SETCC) {
- // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
+ // For vectors:
+ // aext(setcc) -> vsetcc
+ // aext(setcc) -> truncate(vsetcc)
+ // aext(setcc) -> aext(vsetcc)
// Only do this before legalize for now.
if (VT.isVector() && !LegalOperations) {
EVT N0VT = N0.getOperand(0).getValueType();
cast<CondCodeSDNode>(N0.getOperand(2))->get());
// If the desired elements are smaller or larger than the source
// elements we can use a matching integer vector type and then
- // truncate/sign extend
+ // truncate/any extend
else {
- EVT MatchingElementType =
- EVT::getIntegerVT(*DAG.getContext(),
- N0VT.getScalarType().getSizeInBits());
- EVT MatchingVectorType =
- EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
- N0VT.getVectorNumElements());
+ EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
SDValue VsetCC =
DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
N0.getOperand(1),
cast<CondCodeSDNode>(N0.getOperand(2))->get());
- return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
+ return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
}
}
default: break;
case ISD::Constant: {
const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
- assert(CV != 0 && "Const value should be ConstSDNode.");
+ assert(CV && "Const value should be ConstSDNode.");
const APInt &CVal = CV->getAPIntValue();
APInt NewVal = CVal & Mask;
if (NewVal != CVal)
if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
N0.getOperand(1), false);
- if (BSwap.getNode() != 0)
+ if (BSwap.getNode())
return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
BSwap, N1);
}
}
ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
- const APInt &C = CurrentND->getAPIntValue();
- Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt),
+ const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
+ Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
Op.getValueType()));
}
- return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, &Elts[0], NumElts);
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
}
return SDValue();
// creates this pattern) and before operation legalization after which
// we need to be more careful about the vector instructions that we generate.
if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
- LegalTypes && !LegalOperations && N0->hasOneUse()) {
+ LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
EVT VecTy = N0.getOperand(0).getValueType();
EVT ExTy = N0.getValueType();
for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
Opnds.push_back(BuildVect.getOperand(i));
- return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, &Opnds[0],
- Opnds.size());
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
}
}
AddToWorkList(NV.getNode());
Opnds.push_back(NV);
}
- return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
- &Opnds[0], Opnds.size());
+ return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
}
}
LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
- LD1->getPointerInfo().getAddrSpace() !=
- LD2->getPointerInfo().getAddrSpace())
+ LD1->getAddressSpace() != LD2->getAddressSpace())
return SDValue();
EVT LD1VT = LD1->getValueType(0);
if (!LegalTypes &&
N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
VT.isVector()) {
- bool isSimple = true;
- for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
- if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
- N0.getOperand(i).getOpcode() != ISD::Constant &&
- N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
- isSimple = false;
- break;
- }
+ bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
EVT DestEltVT = N->getValueType(0).getVectorElementType();
assert(!DestEltVT.isVector() &&
DstEltVT, Op));
AddToWorkList(Ops.back().getNode());
}
- return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
- &Ops[0], Ops.size());
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
}
// Otherwise, we're growing or shrinking the elements. To avoid having to
}
EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
- return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
- &Ops[0], Ops.size());
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
}
// Finally, this must be the case where we are shrinking elements: each input
std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
}
- return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
- &Ops[0], Ops.size());
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
}
SDValue DAGCombiner::visitFADD(SDNode *N) {
return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
}
- // The next optimizations are desireable only if SELECT_CC can be lowered.
+ // The next optimizations are desirable only if SELECT_CC can be lowered.
// Check against MVT::Other for SELECT_CC, which is a workaround for targets
// having to say they don't support SELECT_CC on every type the DAG knows
// about, since there is no way to mark an opcode illegal at all value types
{ N0.getOperand(0), N0.getOperand(1),
DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
N0.getOperand(2) };
- return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
+ return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
}
// fold (sint_to_fp (zext (setcc x, y, cc))) ->
{ N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
N0.getOperand(0).getOperand(2) };
- return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
+ return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
}
}
return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
}
- // The next optimizations are desireable only if SELECT_CC can be lowered.
+ // The next optimizations are desirable only if SELECT_CC can be lowered.
// Check against MVT::Other for SELECT_CC, which is a workaround for targets
// having to say they don't support SELECT_CC on every type the DAG knows
// about, since there is no way to mark an opcode illegal at all value types
{ N0.getOperand(0), N0.getOperand(1),
DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
N0.getOperand(2) };
- return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
+ return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
}
}
// (fneg (fmul c, x)) -> (fmul -c, x)
if (N0.getOpcode() == ISD::FMUL) {
ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
- if (CFP1)
- return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
- N0.getOperand(0),
- DAG.getNode(ISD::FNEG, SDLoc(N), VT,
- N0.getOperand(1)));
+ if (CFP1) {
+ APFloat CVal = CFP1->getValueAPF();
+ CVal.changeSign();
+ if (Level >= AfterLegalizeDAG &&
+ (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
+ TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
+ return DAG.getNode(
+ ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
+ DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
+ }
}
return SDValue();
((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
(N1.getOperand(0).hasOneUse() &&
N1.getOperand(0).getOpcode() == ISD::SRL))) {
- SDNode *Trunc = 0;
+ SDNode *Trunc = nullptr;
if (N1.getOpcode() == ISD::TRUNCATE) {
// Look pass the truncate.
Trunc = N1.getNode();
// a copy of the original base pointer.
SmallVector<SDNode *, 16> OtherUses;
if (isa<ConstantSDNode>(Offset))
- for (SDNode::use_iterator I = BasePtr.getNode()->use_begin(),
- E = BasePtr.getNode()->use_end(); I != E; ++I) {
- SDNode *Use = *I;
+ for (SDNode *Use : BasePtr.getNode()->uses()) {
if (Use == Ptr.getNode())
continue;
SmallPtrSet<const SDNode *, 32> Visited;
SmallVector<const SDNode *, 16> Worklist;
- for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
- E = Ptr.getNode()->use_end(); I != E; ++I) {
- SDNode *Use = *I;
+ for (SDNode *Use : Ptr.getNode()->uses()) {
if (Use == N)
continue;
if (N->hasPredecessorHelper(Use, Visited, Worklist))
if (Ptr.getNode()->hasOneUse())
return false;
- for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
- E = Ptr.getNode()->use_end(); I != E; ++I) {
- SDNode *Op = *I;
+ for (SDNode *Op : Ptr.getNode()->uses()) {
if (Op == N ||
(Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
continue;
// Check for #1.
bool TryNext = false;
- for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
- EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
- SDNode *Use = *II;
+ for (SDNode *Use : BasePtr.getNode()->uses()) {
if (Use == Ptr.getNode())
continue;
// transformation.
if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
bool RealUse = false;
- for (SDNode::use_iterator III = Use->use_begin(),
- EEE = Use->use_end(); III != EEE; ++III) {
- SDNode *UseUse = *III;
+ for (SDNode *UseUse : Use->uses()) {
if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
RealUse = true;
}
bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
- if (UseAA) {
+#ifndef NDEBUG
+ if (CombinerAAOnlyFunc.getNumOccurrences() &&
+ CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
+ UseAA = false;
+#endif
+ if (UseAA && LD->isUnindexed()) {
// Walk up chain skipping non-aliasing memory nodes.
SDValue BetterChain = FindBetterChain(N, Chain);
// This is used to get some contextual information about legal types, etc.
SelectionDAG *DAG;
- LoadedSlice(SDNode *Inst = NULL, LoadSDNode *Origin = NULL,
- unsigned Shift = 0, SelectionDAG *DAG = NULL)
+ LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
+ unsigned Shift = 0, SelectionDAG *DAG = nullptr)
: Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
LoadedSlice(const LoadedSlice &LS)
/// \brief Get the offset in bytes of this slice in the original chunk of
/// bits.
- /// \pre DAG != NULL.
+ /// \pre DAG != nullptr.
uint64_t getOffsetFromBase() const {
assert(DAG && "Missing context.");
bool IsBigEndian =
};
}
-/// \brief Sorts LoadedSlice according to their offset.
-struct LoadedSliceSorter {
- bool operator()(const LoadedSlice &LHS, const LoadedSlice &RHS) {
- assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
- return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
- }
-};
-
/// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
/// \p UsedBits looks like 0..0 1..1 0..0.
static bool areUsedBitsDense(const APInt &UsedBits) {
// Sort the slices so that elements that are likely to be next to each
// other in memory are next to each other in the list.
- std::sort(LoadedSlices.begin(), LoadedSlices.end(), LoadedSliceSorter());
+ std::sort(LoadedSlices.begin(), LoadedSlices.end(),
+ [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
+ assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
+ return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
+ });
const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
// First (resp. Second) is the first (resp. Second) potentially candidate
// to be placed in a paired load.
- const LoadedSlice *First = NULL;
- const LoadedSlice *Second = NULL;
+ const LoadedSlice *First = nullptr;
+ const LoadedSlice *Second = nullptr;
for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
// Set the beginning of the pair.
First = Second) {
unsigned RequiredAlignment = 0;
if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
// move to the next pair, this type is hopeless.
- Second = NULL;
+ Second = nullptr;
continue;
}
// Check if we meet the alignment requirement.
assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
--GlobalLSCost.Loads;
// Move to the next pair.
- Second = NULL;
+ Second = nullptr;
}
}
// The width of the type must be a power of 2 and greater than 8-bits.
// Otherwise the load cannot be represented in LLVM IR.
// Moreover, if we shifted with a non-8-bits multiple, the slice
- // will be accross several bytes. We do not support that.
+ // will be across several bytes. We do not support that.
unsigned Width = User->getValueSizeInBits(0);
if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
return 0;
}
SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
- &ArgChains[0], ArgChains.size());
+ ArgChains);
DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
return true;
}
// that uses this. If not, this is not a replacement.
APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
ByteShift*8, (ByteShift+NumBytes)*8);
- if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
+ if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
// Check that it is legal on the target to do this. It is legal if the new
// VT we're shrinking to (i8/i16/i32) is legal or we're still before type
// legalization.
MVT VT = MVT::getIntegerVT(NumBytes*8);
if (!DC->isTypeLegal(VT))
- return 0;
+ return nullptr;
// Okay, we can do this! Replace the 'St' store with a store of IVal that is
// shifted by ByteShift and truncated down to NumBytes.
unsigned SequenceNum;
};
-/// Sorts store nodes in a link according to their offset from a shared
-// base ptr.
-struct ConsecutiveMemoryChainSorter {
- bool operator()(MemOpLink LHS, MemOpLink RHS) {
- return LHS.OffsetFromBase < RHS.OffsetFromBase;
- }
-};
-
bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
EVT MemVT = St->getMemoryVT();
int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
break;
} else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
if (Ldn->isVolatile()) {
- Index = NULL;
+ Index = nullptr;
break;
}
NextInChain = Ldn->getChain().getNode();
continue;
} else {
- Index = NULL;
+ Index = nullptr;
break;
}
}
// Sort the memory operands according to their distance from the base pointer.
std::sort(StoreNodes.begin(), StoreNodes.end(),
- ConsecutiveMemoryChainSorter());
+ [](MemOpLink LHS, MemOpLink RHS) {
+ return LHS.OffsetFromBase < RHS.OffsetFromBase ||
+ (LHS.OffsetFromBase == RHS.OffsetFromBase &&
+ LHS.SequenceNum > RHS.SequenceNum);
+ });
// Scan the memory operations on the chain and find the first non-consecutive
// store memory address.
bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
- if (UseAA) {
+#ifndef NDEBUG
+ if (CombinerAAOnlyFunc.getNumOccurrences() &&
+ CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
+ UseAA = false;
+#endif
+ if (UseAA && ST->isUnindexed()) {
// Walk up chain skipping non-aliasing memory nodes.
SDValue BetterChain = FindBetterChain(N, Chain);
}
// Return the new vector
- return DAG.getNode(ISD::BUILD_VECTOR, dl,
- VT, &Ops[0], Ops.size());
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
+}
+
+SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
+ SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
+ EVT ResultVT = EVE->getValueType(0);
+ EVT VecEltVT = InVecVT.getVectorElementType();
+ unsigned Align = OriginalLoad->getAlignment();
+ unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
+ VecEltVT.getTypeForEVT(*DAG.getContext()));
+
+ if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
+ return SDValue();
+
+ Align = NewAlign;
+
+ SDValue NewPtr = OriginalLoad->getBasePtr();
+ SDValue Offset;
+ EVT PtrType = NewPtr.getValueType();
+ MachinePointerInfo MPI;
+ if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
+ int Elt = ConstEltNo->getZExtValue();
+ unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
+ if (TLI.isBigEndian())
+ PtrOff = InVecVT.getSizeInBits() / 8 - PtrOff;
+ Offset = DAG.getConstant(PtrOff, PtrType);
+ MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
+ } else {
+ Offset = DAG.getNode(
+ ISD::MUL, SDLoc(EVE), EltNo.getValueType(), EltNo,
+ DAG.getConstant(VecEltVT.getStoreSize(), EltNo.getValueType()));
+ if (TLI.isBigEndian())
+ Offset = DAG.getNode(
+ ISD::SUB, SDLoc(EVE), EltNo.getValueType(),
+ DAG.getConstant(InVecVT.getStoreSize(), EltNo.getValueType()), Offset);
+ MPI = OriginalLoad->getPointerInfo();
+ }
+ NewPtr = DAG.getNode(ISD::ADD, SDLoc(EVE), PtrType, NewPtr, Offset);
+
+ // The replacement we need to do here is a little tricky: we need to
+ // replace an extractelement of a load with a load.
+ // Use ReplaceAllUsesOfValuesWith to do the replacement.
+ // Note that this replacement assumes that the extractvalue is the only
+ // use of the load; that's okay because we don't want to perform this
+ // transformation in other cases anyway.
+ SDValue Load;
+ SDValue Chain;
+ if (ResultVT.bitsGT(VecEltVT)) {
+ // If the result type of vextract is wider than the load, then issue an
+ // extending load instead.
+ ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, VecEltVT)
+ ? ISD::ZEXTLOAD
+ : ISD::EXTLOAD;
+ Load = DAG.getExtLoad(ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(),
+ NewPtr, MPI, VecEltVT, OriginalLoad->isVolatile(),
+ OriginalLoad->isNonTemporal(), Align,
+ OriginalLoad->getTBAAInfo());
+ Chain = Load.getValue(1);
+ } else {
+ Load = DAG.getLoad(
+ VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
+ OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
+ OriginalLoad->isInvariant(), Align, OriginalLoad->getTBAAInfo());
+ Chain = Load.getValue(1);
+ if (ResultVT.bitsLT(VecEltVT))
+ Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
+ else
+ Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
+ }
+ WorkListRemover DeadNodes(*this);
+ SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
+ SDValue To[] = { Load, Chain };
+ DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
+ // Since we're explicitly calling ReplaceAllUses, add the new node to the
+ // worklist explicitly as well.
+ AddToWorkList(Load.getNode());
+ AddUsersToWorkList(Load.getNode()); // Add users too
+ // Make sure to revisit this node to clean it up; it will usually be dead.
+ AddToWorkList(EVE);
+ ++OpsNarrowed;
+ return SDValue(EVE, 0);
}
SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
// We only perform this optimization before the op legalization phase because
// we may introduce new vector instructions which are not backed by TD
// patterns. For example on AVX, extracting elements from a wide vector
- // without using extract_subvector.
+ // without using extract_subvector. However, if we can find an underlying
+ // scalar value, then we can always use that.
if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
- && ConstEltNo && !LegalOperations) {
+ && ConstEltNo) {
int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
int NumElem = VT.getVectorNumElements();
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
return DAG.getUNDEF(NVT);
// Select the right vector half to extract from.
+ SDValue SVInVec;
if (OrigElt < NumElem) {
- InVec = InVec->getOperand(0);
+ SVInVec = InVec->getOperand(0);
} else {
- InVec = InVec->getOperand(1);
+ SVInVec = InVec->getOperand(1);
OrigElt -= NumElem;
}
- EVT IndexTy = TLI.getVectorIdxTy();
- return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
- InVec, DAG.getConstant(OrigElt, IndexTy));
+ if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
+ SDValue InOp = SVInVec.getOperand(OrigElt);
+ if (InOp.getValueType() != NVT) {
+ assert(InOp.getValueType().isInteger() && NVT.isInteger());
+ InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
+ }
+
+ return InOp;
+ }
+
+ // FIXME: We should handle recursing on other vector shuffles and
+ // scalar_to_vector here as well.
+
+ if (!LegalOperations) {
+ EVT IndexTy = TLI.getVectorIdxTy();
+ return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
+ SVInVec, DAG.getConstant(OrigElt, IndexTy));
+ }
+ }
+
+ bool BCNumEltsChanged = false;
+ EVT ExtVT = VT.getVectorElementType();
+ EVT LVT = ExtVT;
+
+ // If the result of load has to be truncated, then it's not necessarily
+ // profitable.
+ if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
+ return SDValue();
+
+ if (InVec.getOpcode() == ISD::BITCAST) {
+ // Don't duplicate a load with other uses.
+ if (!InVec.hasOneUse())
+ return SDValue();
+
+ EVT BCVT = InVec.getOperand(0).getValueType();
+ if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
+ return SDValue();
+ if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
+ BCNumEltsChanged = true;
+ InVec = InVec.getOperand(0);
+ ExtVT = BCVT.getVectorElementType();
+ }
+
+ // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
+ if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
+ ISD::isNormalLoad(InVec.getNode())) {
+ SDValue Index = N->getOperand(1);
+ if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
+ return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
+ OrigLoad);
}
// Perform only after legalization to ensure build_vector / vector_shuffle
if (ConstEltNo) {
int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
- bool NewLoad = false;
- bool BCNumEltsChanged = false;
- EVT ExtVT = VT.getVectorElementType();
- EVT LVT = ExtVT;
-
- // If the result of load has to be truncated, then it's not necessarily
- // profitable.
- if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
- return SDValue();
-
- if (InVec.getOpcode() == ISD::BITCAST) {
- // Don't duplicate a load with other uses.
- if (!InVec.hasOneUse())
- return SDValue();
- EVT BCVT = InVec.getOperand(0).getValueType();
- if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
- return SDValue();
- if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
- BCNumEltsChanged = true;
- InVec = InVec.getOperand(0);
- ExtVT = BCVT.getVectorElementType();
- NewLoad = true;
- }
-
- LoadSDNode *LN0 = NULL;
- const ShuffleVectorSDNode *SVN = NULL;
+ LoadSDNode *LN0 = nullptr;
+ const ShuffleVectorSDNode *SVN = nullptr;
if (ISD::isNormalLoad(InVec.getNode())) {
LN0 = cast<LoadSDNode>(InVec);
} else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
if (ISD::isNormalLoad(InVec.getNode())) {
LN0 = cast<LoadSDNode>(InVec);
Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
+ EltNo = DAG.getConstant(Elt, EltNo.getValueType());
}
}
if (Elt == -1)
return DAG.getUNDEF(LVT);
- unsigned Align = LN0->getAlignment();
- if (NewLoad) {
- // Check the resultant load doesn't need a higher alignment than the
- // original load.
- unsigned NewAlign =
- TLI.getDataLayout()
- ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
-
- if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
- return SDValue();
-
- Align = NewAlign;
- }
-
- SDValue NewPtr = LN0->getBasePtr();
- unsigned PtrOff = 0;
-
- if (Elt) {
- PtrOff = LVT.getSizeInBits() * Elt / 8;
- EVT PtrType = NewPtr.getValueType();
- if (TLI.isBigEndian())
- PtrOff = VT.getSizeInBits() / 8 - PtrOff;
- NewPtr = DAG.getNode(ISD::ADD, SDLoc(N), PtrType, NewPtr,
- DAG.getConstant(PtrOff, PtrType));
- }
-
- // The replacement we need to do here is a little tricky: we need to
- // replace an extractelement of a load with a load.
- // Use ReplaceAllUsesOfValuesWith to do the replacement.
- // Note that this replacement assumes that the extractvalue is the only
- // use of the load; that's okay because we don't want to perform this
- // transformation in other cases anyway.
- SDValue Load;
- SDValue Chain;
- if (NVT.bitsGT(LVT)) {
- // If the result type of vextract is wider than the load, then issue an
- // extending load instead.
- ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
- ? ISD::ZEXTLOAD : ISD::EXTLOAD;
- Load = DAG.getExtLoad(ExtType, SDLoc(N), NVT, LN0->getChain(),
- NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
- LVT, LN0->isVolatile(), LN0->isNonTemporal(),
- Align, LN0->getTBAAInfo());
- Chain = Load.getValue(1);
- } else {
- Load = DAG.getLoad(LVT, SDLoc(N), LN0->getChain(), NewPtr,
- LN0->getPointerInfo().getWithOffset(PtrOff),
- LN0->isVolatile(), LN0->isNonTemporal(),
- LN0->isInvariant(), Align, LN0->getTBAAInfo());
- Chain = Load.getValue(1);
- if (NVT.bitsLT(LVT))
- Load = DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, Load);
- else
- Load = DAG.getNode(ISD::BITCAST, SDLoc(N), NVT, Load);
- }
- WorkListRemover DeadNodes(*this);
- SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
- SDValue To[] = { Load, Chain };
- DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
- // Since we're explcitly calling ReplaceAllUses, add the new node to the
- // worklist explicitly as well.
- AddToWorkList(Load.getNode());
- AddUsersToWorkList(Load.getNode()); // Add users too
- // Make sure to revisit this node to clean it up; it will usually be dead.
- AddToWorkList(N);
- return SDValue(N, 0);
+ return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
}
return SDValue();
if (!isTypeLegal(VecVT)) return SDValue();
// Make the new BUILD_VECTOR.
- SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], Ops.size());
+ SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
// The new BUILD_VECTOR node has the potential to be further optimized.
AddToWorkList(BV.getNode());
else
Opnds.push_back(In.getOperand(0));
}
- SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT,
- &Opnds[0], Opnds.size());
+ SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
AddToWorkList(BV.getNode());
return DAG.getNode(Opcode, dl, VT, BV);
// constant index, bail out.
if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
!isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
- VecIn1 = VecIn2 = SDValue(0, 0);
+ VecIn1 = VecIn2 = SDValue(nullptr, 0);
break;
}
if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
continue;
- if (VecIn1.getNode() == 0) {
+ if (!VecIn1.getNode()) {
VecIn1 = ExtractedFromVec;
- } else if (VecIn2.getNode() == 0) {
+ } else if (!VecIn2.getNode()) {
VecIn2 = ExtractedFromVec;
} else {
// Too many inputs.
- VecIn1 = VecIn2 = SDValue(0, 0);
+ VecIn1 = VecIn2 = SDValue(nullptr, 0);
break;
}
}
- // If everything is good, we can make a shuffle operation.
+ // If everything is good, we can make a shuffle operation.
if (VecIn1.getNode()) {
SmallVector<int, 8> Mask;
for (unsigned i = 0; i != NumInScalars; ++i) {
// Attempt to transform a single input vector to the correct type.
if ((VT != VecIn1.getValueType())) {
// We don't support shuffeling between TWO values of different types.
- if (VecIn2.getNode() != 0)
+ if (VecIn2.getNode())
return SDValue();
// We only support widening of vectors which are half the size of the
}
}
+ // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
+ // -> (BUILD_VECTOR A, B, ..., C, D, ...)
+ if (N->getNumOperands() == 2 &&
+ N->getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
+ N->getOperand(1).getOpcode() == ISD::BUILD_VECTOR) {
+ EVT VT = N->getValueType(0);
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ SmallVector<SDValue, 8> Opnds;
+ unsigned BuildVecNumElts = N0.getNumOperands();
+
+ for (unsigned i = 0; i != BuildVecNumElts; ++i)
+ Opnds.push_back(N0.getOperand(i));
+ for (unsigned i = 0; i != BuildVecNumElts; ++i)
+ Opnds.push_back(N1.getOperand(i));
+
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
+ }
+
// Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
// nodes often generate nop CONCAT_VECTOR nodes.
// Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
}
}
- return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops.data(),
- Ops.size());
+ return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
}
SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
return SDValue();
}
+SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
+ SDValue N0 = N->getOperand(0);
+ SDValue N2 = N->getOperand(2);
+
+ // If the input vector is a concatenation, and the insert replaces
+ // one of the halves, we can optimize into a single concat_vectors.
+ if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
+ N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
+ APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
+ EVT VT = N->getValueType(0);
+
+ // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
+ // (concat_vectors Z, Y)
+ if (InsIdx == 0)
+ return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
+ N->getOperand(1), N0.getOperand(1));
+
+ // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
+ // (concat_vectors X, Z)
+ if (InsIdx == VT.getVectorNumElements()/2)
+ return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
+ N0.getOperand(0), N->getOperand(1));
+ }
+
+ return SDValue();
+}
+
/// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
/// an AND to a vector_shuffle with the destination vector and a zero vector.
/// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
EVT EltVT = RVT.getVectorElementType();
SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
DAG.getConstant(0, EltVT));
- SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
- RVT, &ZeroOps[0], ZeroOps.size());
+ SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
// this operation.
if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
RHS.getOpcode() == ISD::BUILD_VECTOR) {
+ // Check if both vectors are constants. If not bail out.
+ if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
+ cast<BuildVectorSDNode>(RHS)->isConstant()))
+ return SDValue();
+
SmallVector<SDValue, 8> Ops;
for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
SDValue LHSOp = LHS.getOperand(i);
SDValue RHSOp = RHS.getOperand(i);
- // If these two elements can't be folded, bail out.
- if ((LHSOp.getOpcode() != ISD::UNDEF &&
- LHSOp.getOpcode() != ISD::Constant &&
- LHSOp.getOpcode() != ISD::ConstantFP) ||
- (RHSOp.getOpcode() != ISD::UNDEF &&
- RHSOp.getOpcode() != ISD::Constant &&
- RHSOp.getOpcode() != ISD::ConstantFP))
- break;
// Can't fold divide by zero.
if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
}
if (Ops.size() == LHS.getNumOperands())
- return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
- LHS.getValueType(), &Ops[0], Ops.size());
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
+ }
+
+ // Type legalization might introduce new shuffles in the DAG.
+ // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
+ // -> (shuffle (VBinOp (A, B)), Undef, Mask).
+ if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
+ isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
+ LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
+ RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
+ ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
+ ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
+
+ if (SVN0->getMask().equals(SVN1->getMask())) {
+ EVT VT = N->getValueType(0);
+ SDValue UndefVector = LHS.getOperand(1);
+ SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
+ LHS.getOperand(0), RHS.getOperand(0));
+ AddUsersToWorkList(N);
+ return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
+ &SVN0->getMask()[0]);
+ }
}
return SDValue();
if (Ops.size() != N0.getNumOperands())
return SDValue();
- return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
- N0.getValueType(), &Ops[0], Ops.size());
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N0.getValueType(), Ops);
}
SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
if (TLI.isTypeLegal(N2.getValueType()) &&
(TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
- TargetLowering::Legal) &&
+ TargetLowering::Legal &&
+ !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
+ !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
// If both constants have multiple uses, then we won't need to do an
// extra load, they are likely around in registers for other users.
(TV->hasOneUse() || FV->hasOneUse())) {
// select_cc setlt X, 1, -X, X ->
// Y = sra (X, size(X)-1); xor (add (X, Y), Y)
if (N1C) {
- ConstantSDNode *SubC = NULL;
+ ConstantSDNode *SubC = nullptr;
if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
(N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
/// multiplying by a magic number. See:
/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
SDValue DAGCombiner::BuildSDIV(SDNode *N) {
+ ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
+ if (!C)
+ return SDValue();
+
+ // Avoid division by zero.
+ if (!C->getAPIntValue())
+ return SDValue();
+
std::vector<SDNode*> Built;
- SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
+ SDValue S =
+ TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
- for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
- ii != ee; ++ii)
- AddToWorkList(*ii);
+ for (SDNode *N : Built)
+ AddToWorkList(N);
return S;
}
-/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
+/// BuildUDIV - Given an ISD::UDIV node expressing a divide by constant,
/// return a DAG expression to select that will generate the same value by
/// multiplying by a magic number. See:
/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
SDValue DAGCombiner::BuildUDIV(SDNode *N) {
+ ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
+ if (!C)
+ return SDValue();
+
+ // Avoid division by zero.
+ if (!C->getAPIntValue())
+ return SDValue();
+
std::vector<SDNode*> Built;
- SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
+ SDValue S =
+ TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
- for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
- ii != ee; ++ii)
- AddToWorkList(*ii);
+ for (SDNode *N : Built)
+ AddToWorkList(N);
return S;
}
static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
const GlobalValue *&GV, const void *&CV) {
// Assume it is a primitive operation.
- Base = Ptr; Offset = 0; GV = 0; CV = 0;
+ Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
// If it's an adding a simple constant then integrate the offset.
if (Base.getOpcode() == ISD::ADD) {
/// isAlias - Return true if there is any possibility that the two addresses
/// overlap.
-bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1, bool IsVolatile1,
- const Value *SrcValue1, int SrcValueOffset1,
- unsigned SrcValueAlign1,
- const MDNode *TBAAInfo1,
- SDValue Ptr2, int64_t Size2, bool IsVolatile2,
- const Value *SrcValue2, int SrcValueOffset2,
- unsigned SrcValueAlign2,
- const MDNode *TBAAInfo2) const {
+bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
// If they are the same then they must be aliases.
- if (Ptr1 == Ptr2) return true;
+ if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
// If they are both volatile then they cannot be reordered.
- if (IsVolatile1 && IsVolatile2) return true;
+ if (Op0->isVolatile() && Op1->isVolatile()) return true;
// Gather base node and offset information.
SDValue Base1, Base2;
int64_t Offset1, Offset2;
const GlobalValue *GV1, *GV2;
const void *CV1, *CV2;
- bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
- bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
+ bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
+ Base1, Offset1, GV1, CV1);
+ bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
+ Base2, Offset2, GV2, CV2);
// If they have a same base address then check to see if they overlap.
if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
- return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
+ return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
+ (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
// It is possible for different frame indices to alias each other, mostly
// when tail call optimization reuses return address slots for arguments.
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
- return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
+ return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
+ (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
}
// Otherwise, if we know what the bases are, and they aren't identical, then
// compared to the size and offset of the access, we may be able to prove they
// do not alias. This check is conservative for now to catch cases created by
// splitting vector types.
- if ((SrcValueAlign1 == SrcValueAlign2) &&
- (SrcValueOffset1 != SrcValueOffset2) &&
- (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
- int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
- int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
+ if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
+ (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
+ (Op0->getMemoryVT().getSizeInBits() >> 3 ==
+ Op1->getMemoryVT().getSizeInBits() >> 3) &&
+ (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
+ int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
+ int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
// There is no overlap between these relatively aligned accesses of similar
// size, return no alias.
- if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
+ if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
+ (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
return false;
}
bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA :
TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
- if (UseAA && SrcValue1 && SrcValue2) {
+#ifndef NDEBUG
+ if (CombinerAAOnlyFunc.getNumOccurrences() &&
+ CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
+ UseAA = false;
+#endif
+ if (UseAA &&
+ Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
// Use alias analysis information.
- int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
- int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
- int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
+ int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
+ Op1->getSrcValueOffset());
+ int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
+ Op0->getSrcValueOffset() - MinOffset;
+ int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
+ Op1->getSrcValueOffset() - MinOffset;
AliasAnalysis::AliasResult AAResult =
- AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
- AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
+ AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
+ Overlap1,
+ UseTBAA ? Op0->getTBAAInfo() : nullptr),
+ AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
+ Overlap2,
+ UseTBAA ? Op1->getTBAAInfo() : nullptr));
if (AAResult == AliasAnalysis::NoAlias)
return false;
}
return true;
}
-bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) {
- SDValue Ptr0, Ptr1;
- int64_t Size0, Size1;
- bool IsVolatile0, IsVolatile1;
- const Value *SrcValue0, *SrcValue1;
- int SrcValueOffset0, SrcValueOffset1;
- unsigned SrcValueAlign0, SrcValueAlign1;
- const MDNode *SrcTBAAInfo0, *SrcTBAAInfo1;
- FindAliasInfo(Op0, Ptr0, Size0, IsVolatile0, SrcValue0, SrcValueOffset0,
- SrcValueAlign0, SrcTBAAInfo0);
- FindAliasInfo(Op1, Ptr1, Size1, IsVolatile1, SrcValue1, SrcValueOffset1,
- SrcValueAlign1, SrcTBAAInfo1);
- return isAlias(Ptr0, Size0, IsVolatile0, SrcValue0, SrcValueOffset0,
- SrcValueAlign0, SrcTBAAInfo0,
- Ptr1, Size1, IsVolatile1, SrcValue1, SrcValueOffset1,
- SrcValueAlign1, SrcTBAAInfo1);
-}
-
-/// FindAliasInfo - Extracts the relevant alias information from the memory
-/// node. Returns true if the operand was a nonvolatile load.
-bool DAGCombiner::FindAliasInfo(SDNode *N,
- SDValue &Ptr, int64_t &Size, bool &IsVolatile,
- const Value *&SrcValue,
- int &SrcValueOffset,
- unsigned &SrcValueAlign,
- const MDNode *&TBAAInfo) const {
- LSBaseSDNode *LS = cast<LSBaseSDNode>(N);
-
- Ptr = LS->getBasePtr();
- Size = LS->getMemoryVT().getSizeInBits() >> 3;
- IsVolatile = LS->isVolatile();
- SrcValue = LS->getSrcValue();
- SrcValueOffset = LS->getSrcValueOffset();
- SrcValueAlign = LS->getOriginalAlignment();
- TBAAInfo = LS->getTBAAInfo();
- return isa<LoadSDNode>(LS) && !IsVolatile;
-}
-
/// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
/// looking for aliasing nodes and adding them to the Aliases vector.
void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
// Get alias information for node.
- SDValue Ptr;
- int64_t Size;
- bool IsVolatile;
- const Value *SrcValue;
- int SrcValueOffset;
- unsigned SrcValueAlign;
- const MDNode *SrcTBAAInfo;
- bool IsLoad = FindAliasInfo(N, Ptr, Size, IsVolatile, SrcValue,
- SrcValueOffset, SrcValueAlign, SrcTBAAInfo);
+ bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
// Starting off.
Chains.push_back(OriginalChain);
if (Depth > 6 || Aliases.size() == 2) {
Aliases.clear();
Aliases.push_back(OriginalChain);
- break;
+ return;
}
// Don't bother if we've been before.
case ISD::LOAD:
case ISD::STORE: {
// Get alias information for Chain.
- SDValue OpPtr;
- int64_t OpSize;
- bool OpIsVolatile;
- const Value *OpSrcValue;
- int OpSrcValueOffset;
- unsigned OpSrcValueAlign;
- const MDNode *OpSrcTBAAInfo;
- bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
- OpIsVolatile, OpSrcValue, OpSrcValueOffset,
- OpSrcValueAlign,
- OpSrcTBAAInfo);
+ bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
+ !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
// If chain is alias then stop here.
if (!(IsLoad && IsOpLoad) &&
- isAlias(Ptr, Size, IsVolatile, SrcValue, SrcValueOffset,
- SrcValueAlign, SrcTBAAInfo,
- OpPtr, OpSize, OpIsVolatile, OpSrcValue, OpSrcValueOffset,
- OpSrcValueAlign, OpSrcTBAAInfo)) {
+ isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
Aliases.push_back(Chain);
} else {
// Look further up the chain.
break;
}
}
+
+ // We need to be careful here to also search for aliases through the
+ // value operand of a store, etc. Consider the following situation:
+ // Token1 = ...
+ // L1 = load Token1, %52
+ // S1 = store Token1, L1, %51
+ // L2 = load Token1, %52+8
+ // S2 = store Token1, L2, %51+8
+ // Token2 = Token(S1, S2)
+ // L3 = load Token2, %53
+ // S3 = store Token2, L3, %52
+ // L4 = load Token2, %53+8
+ // S4 = store Token2, L4, %52+8
+ // If we search for aliases of S3 (which loads address %52), and we look
+ // only through the chain, then we'll miss the trivial dependence on L1
+ // (which also loads from %52). We then might change all loads and
+ // stores to use Token1 as their chain operand, which could result in
+ // copying %53 into %52 before copying %52 into %51 (which should
+ // happen first).
+ //
+ // The problem is, however, that searching for such data dependencies
+ // can become expensive, and the cost is not directly related to the
+ // chain depth. Instead, we'll rule out such configurations here by
+ // insisting that we've visited all chain users (except for users
+ // of the original chain, which is not necessary). When doing this,
+ // we need to look through nodes we don't care about (otherwise, things
+ // like register copies will interfere with trivial cases).
+
+ SmallVector<const SDNode *, 16> Worklist;
+ for (SmallPtrSet<SDNode *, 16>::iterator I = Visited.begin(),
+ IE = Visited.end(); I != IE; ++I)
+ if (*I != OriginalChain.getNode())
+ Worklist.push_back(*I);
+
+ while (!Worklist.empty()) {
+ const SDNode *M = Worklist.pop_back_val();
+
+ // We have already visited M, and want to make sure we've visited any uses
+ // of M that we care about. For uses that we've not visisted, and don't
+ // care about, queue them to the worklist.
+
+ for (SDNode::use_iterator UI = M->use_begin(),
+ UIE = M->use_end(); UI != UIE; ++UI)
+ if (UI.getUse().getValueType() == MVT::Other && Visited.insert(*UI)) {
+ if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
+ // We've not visited this use, and we care about it (it could have an
+ // ordering dependency with the original node).
+ Aliases.clear();
+ Aliases.push_back(OriginalChain);
+ return;
+ }
+
+ // We've not visited this use, but we don't care about it. Mark it as
+ // visited and enqueue it to the worklist.
+ Worklist.push_back(*UI);
+ }
+ }
}
/// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
return Aliases[0];
// Construct a custom tailored token factor.
- return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
- &Aliases[0], Aliases.size());
+ return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
}
// SelectionDAG::Combine - This is the entry point for the file.