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"));
+
+// FIXME: Enable the use of TBAA. There are two known issues preventing this:
+// 1. Stack coloring does not update TBAA when merging allocas
+// 2. CGP inserts ptrtoint/inttoptr pairs when sinking address computations.
+// Because BasicAA does not handle inttoptr, we'll often miss basic type
+// punning idioms that we need to catch so we don't miscompile real-world
+// code.
+ static cl::opt<bool>
+ UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(false),
+ 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.
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);
/// \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),
return false;
}
+// \brief Returns the SDNode if it is a constant BuildVector or constant int.
+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 NULL;
+}
+
SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
SDValue N0, SDValue N1) {
EVT VT = N0.getValueType();
- if (VT.isVector()) {
- if (N0.getOpcode() == Opc) {
- BuildVectorSDNode *L = dyn_cast<BuildVectorSDNode>(N0.getOperand(1));
- if(L && L->isConstant()) {
- BuildVectorSDNode *R = dyn_cast<BuildVectorSDNode>(N1);
- if (R && R->isConstant()) {
- // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
- SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R);
- 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 (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 (N1.getOpcode() == Opc) {
- BuildVectorSDNode *R = dyn_cast<BuildVectorSDNode>(N1.getOperand(1));
- if (R && R->isConstant()) {
- BuildVectorSDNode *L = dyn_cast<BuildVectorSDNode>(N0);
- if (L && L->isConstant()) {
- // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
- SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L);
- 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.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));
}
}
-
- return SDValue();
}
- 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 (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));
+ }
}
}
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();
}
// 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);
+ }
}
}
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());
return CombineTo(N, Res, Res);
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
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)){
/// visitShiftByConstant - Handle transforms common to the three shifts, when
/// the shift amount is a constant.
SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
+ assert(isa<ConstantSDNode>(N->getOperand(1)) &&
+ "Expected an ConstantSDNode operand.");
+ // We can't and shouldn't fold opaque constants.
+ if (cast<ConstantSDNode>(N->getOperand(1))->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:
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 = dyn_cast<ConstantSDNode>(N01)) {
+ EVT TruncVT = N->getValueType(0);
+ SDValue N00 = N->getOperand(0).getOperand(0);
+ APInt TruncC = N01C->getAPIntValue();
+ TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
+
+ return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
+ DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
+ DAG.getConstant(TruncC, TruncVT));
+ }
+ }
+
+ return SDValue();
+}
SDValue DAGCombiner::visitSHL(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
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() &&
+ TLI.getBooleanContents(true) ==
+ TargetLowering::ZeroOrNegativeOneBooleanContent &&
+ N0.getOpcode() == ISD::AND) {
+ 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);
+ }
+ }
}
// 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)))
// 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)));
- }
+ 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 (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
// 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
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 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);
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 0;
+
+ // 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[0], NumElts).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
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)
}
}
- // 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));
+
+ }
}
}
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)
if (N0.getOpcode() == ISD::SETCC) {
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));
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)
// 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();
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);
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
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
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);
};
}
-/// \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.
// 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;
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;
// 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);
}
}
+ // 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[0],
+ Opnds.size());
+ }
+
// 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 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>. ==>
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())
+ if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
+ cast<BuildVectorSDNode>(RHS)->isConstant()))
return SDValue();
SmallVector<SDValue, 8> Ops;
bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA :
TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
+#ifndef NDEBUG
+ if (CombinerAAOnlyFunc.getNumOccurrences() &&
+ CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
+ UseAA = false;
+#endif
if (UseAA && SrcValue1 && SrcValue2) {
// Use alias analysis information.
int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
AliasAnalysis::AliasResult AAResult =
- AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
- AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
+ AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1,
+ UseTBAA ? TBAAInfo1 : 0),
+ AliasAnalysis::Location(SrcValue2, Overlap2,
+ UseTBAA ? TBAAInfo2 : 0));
if (AAResult == AliasAnalysis::NoAlias)
return false;
}
if (Depth > 6 || Aliases.size() == 2) {
Aliases.clear();
Aliases.push_back(OriginalChain);
- break;
+ return;
}
// Don't bother if we've been before.
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
projects / oota-llvm.git / blobdiff