[AArch64] Improve lowering for scalable masked interleaving stores

llvm/include/llvm/CodeGen/SelectionDAGNodes.h

@@ -3346,6 +3346,14 @@ namespace ISD {
            Ld->getAddressingMode() == ISD::UNINDEXED;
   }
 
+  /// Returns true if the specified node is a non-extending and unindexed
+  /// masked store.
+  inline bool isNormalMaskedStore(const SDNode *N) {
+    auto *St = dyn_cast<MaskedStoreSDNode>(N);
+    return St && !St->isTruncatingStore() &&
+           St->getAddressingMode() == ISD::UNINDEXED;
+  }
+
   /// Attempt to match a unary predicate against a scalar/splat constant or
   /// every element of a constant BUILD_VECTOR.
   /// If AllowUndef is true, then UNDEF elements will pass nullptr to Match.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -24632,6 +24632,105 @@ static SDValue performSTORECombine(SDNode *N,
   return SDValue();
 }
 
+static bool
+isSequentialConcatOfVectorInterleave(SDNode *N, SmallVectorImpl<SDValue> &Ops) {
+  if (N->getOpcode() != ISD::CONCAT_VECTORS)
+    return false;
+
+  unsigned NumParts = N->getNumOperands();
+
+  // We should be concatenating each sequential result from a
+  // VECTOR_INTERLEAVE.
+  SDNode *InterleaveOp = N->getOperand(0).getNode();
+  if (InterleaveOp->getOpcode() != ISD::VECTOR_INTERLEAVE ||
+      InterleaveOp->getNumOperands() != NumParts)
+    return false;
+
+  for (unsigned I = 0; I < NumParts; I++)
+    if (N->getOperand(I) != SDValue(InterleaveOp, I))
+      return false;
+
+  Ops.append(InterleaveOp->op_begin(), InterleaveOp->op_end());
+  return true;
+}
+
+static SDValue getNarrowMaskForInterleavedOps(SelectionDAG &DAG, SDLoc &DL,
+                                              SDValue WideMask,
+                                              unsigned RequiredNumParts) {
+  if (WideMask->getOpcode() == ISD::CONCAT_VECTORS) {
+    SmallVector<SDValue, 4> MaskInterleaveOps;
+    if (!isSequentialConcatOfVectorInterleave(WideMask.getNode(),
+                                              MaskInterleaveOps))
+      return SDValue();
+
+    if (MaskInterleaveOps.size() != RequiredNumParts)
+      return SDValue();
+
+    // Make sure the inputs to the vector interleave are identical.
+    if (!llvm::all_equal(MaskInterleaveOps))
+      return SDValue();
+
+    return MaskInterleaveOps[0];
+  }
+
+  if (WideMask->getOpcode() != ISD::SPLAT_VECTOR)
+    return SDValue();
+
+  ElementCount EC = WideMask.getValueType().getVectorElementCount();
+  assert(EC.isKnownMultipleOf(RequiredNumParts) &&
+         "Expected element count divisible by number of parts");
+  EC = EC.divideCoefficientBy(RequiredNumParts);
+  return DAG.getNode(ISD::SPLAT_VECTOR, DL, MVT::getVectorVT(MVT::i1, EC),
+                     WideMask->getOperand(0));
+}
+
+static SDValue performInterleavedMaskedStoreCombine(
+    SDNode *N, TargetLowering::DAGCombinerInfo &DCI, SelectionDAG &DAG) {
+  if (!DCI.isBeforeLegalize())
+    return SDValue();
+
+  MaskedStoreSDNode *MST = cast<MaskedStoreSDNode>(N);
+  SDValue WideValue = MST->getValue();
+
+  // Bail out if the stored value has an unexpected number of uses, since we'll
+  // have to perform manual interleaving and may as well just use normal masked
+  // stores. Also, discard masked stores that are truncating or indexed.
+  if (!WideValue.hasOneUse() || !ISD::isNormalMaskedStore(MST) ||
+      !MST->isSimple() || !MST->getOffset().isUndef())
+    return SDValue();
+
+  SmallVector<SDValue, 4> ValueInterleaveOps;
+  if (!isSequentialConcatOfVectorInterleave(WideValue.getNode(),
+                                            ValueInterleaveOps))
+    return SDValue();
+
+  unsigned NumParts = ValueInterleaveOps.size();
+  if (NumParts != 2 && NumParts != 4)
+    return SDValue();
+
+  // At the moment we're unlikely to see a fixed-width vector interleave as
+  // we usually generate shuffles instead.
+  EVT SubVecTy = ValueInterleaveOps[0].getValueType();
+  if (!SubVecTy.isScalableVT() ||
+      SubVecTy.getSizeInBits().getKnownMinValue() != 128 ||
+      !DAG.getTargetLoweringInfo().isTypeLegal(SubVecTy))
+    return SDValue();
+
+  SDLoc DL(N);
+  SDValue NarrowMask =
+      getNarrowMaskForInterleavedOps(DAG, DL, MST->getMask(), NumParts);
+  if (!NarrowMask)
+    return SDValue();
+
+  const Intrinsic::ID IID =
+      NumParts == 2 ? Intrinsic::aarch64_sve_st2 : Intrinsic::aarch64_sve_st4;
+  SmallVector<SDValue, 8> NewStOps;
+  NewStOps.append({MST->getChain(), DAG.getConstant(IID, DL, MVT::i32)});
+  NewStOps.append(ValueInterleaveOps);
+  NewStOps.append({NarrowMask, MST->getBasePtr()});
+  return DAG.getNode(ISD::INTRINSIC_VOID, DL, MVT::Other, NewStOps);
+}
+
 static SDValue performMSTORECombine(SDNode *N,
                                     TargetLowering::DAGCombinerInfo &DCI,
                                     SelectionDAG &DAG,
@@ -24641,6 +24740,9 @@ static SDValue performMSTORECombine(SDNode *N,
   SDValue Mask = MST->getMask();
   SDLoc DL(N);
 
+  if (SDValue Res = performInterleavedMaskedStoreCombine(N, DCI, DAG))
+    return Res;
+
   // If this is a UZP1 followed by a masked store, fold this into a masked
   // truncating store.  We can do this even if this is already a masked
   // truncstore.
@@ -27274,43 +27376,11 @@ static SDValue performVectorDeinterleaveCombine(
     return SDValue();
 
   // Now prove that the mask is an interleave of identical masks.
-  SDValue Mask = MaskedLoad->getMask();
-  if (Mask->getOpcode() != ISD::SPLAT_VECTOR &&
-      Mask->getOpcode() != ISD::CONCAT_VECTORS)
-    return SDValue();
-
-  SDValue NarrowMask;
   SDLoc DL(N);
-  if (Mask->getOpcode() == ISD::CONCAT_VECTORS) {
-    if (Mask->getNumOperands() != NumParts)
-      return SDValue();
-
-    // We should be concatenating each sequential result from a
-    // VECTOR_INTERLEAVE.
-    SDNode *InterleaveOp = Mask->getOperand(0).getNode();
-    if (InterleaveOp->getOpcode() != ISD::VECTOR_INTERLEAVE ||
-        InterleaveOp->getNumOperands() != NumParts)
-      return SDValue();
-
-    for (unsigned I = 0; I < NumParts; I++) {
-      if (Mask.getOperand(I) != SDValue(InterleaveOp, I))
-        return SDValue();
-    }
-
-    // Make sure the inputs to the vector interleave are identical.
-    if (!llvm::all_equal(InterleaveOp->op_values()))
-      return SDValue();
-
-    NarrowMask = InterleaveOp->getOperand(0);
-  } else { // ISD::SPLAT_VECTOR
-    ElementCount EC = Mask.getValueType().getVectorElementCount();
-    assert(EC.isKnownMultipleOf(NumParts) &&
-           "Expected element count divisible by number of parts");
-    EC = EC.divideCoefficientBy(NumParts);
-    NarrowMask =
-        DAG.getNode(ISD::SPLAT_VECTOR, DL, MVT::getVectorVT(MVT::i1, EC),
-                    Mask->getOperand(0));
-  }
+  SDValue NarrowMask =
+      getNarrowMaskForInterleavedOps(DAG, DL, MaskedLoad->getMask(), NumParts);
+  if (!NarrowMask)
+    return SDValue();
 
   const Intrinsic::ID IID = NumParts == 2 ? Intrinsic::aarch64_sve_ld2_sret
                                           : Intrinsic::aarch64_sve_ld4_sret;