[AArch64] Add an all-in-one histogram intrinsic

Based on discussion from
https://discourse.llvm.org/t/rfc-vectorization-support-for-histogram-count-operations/74788

Current interface is:

llvm.experimental.histogram(<vecty> ptrs, <intty> inc_amount, <vecty> mask)

The integer type used by 'inc_amount' needs to match the type of the buckets in memory.

The intrinsic covers the following operations:
  * Gather load
  * histogram on the elements of 'ptrs'
  * multiply the histogram results by 'inc_amount'
  * add the result of the multiply to the values loaded by the gather
  * scatter store the results of the add

Supports lowering to histcnt instructions for AArch64 targets, and scalarization for all others at present.

Author: huntergr-arm

llvm/docs/LangRef.rst

@@ -19143,6 +19143,60 @@ will be on any later loop iteration.
 This intrinsic will only return 0 if the input count is also 0. A non-zero input
 count will produce a non-zero result.
 
+'``llvm.experimental.vector.histogram.*``' Intrinsics
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+These intrinsics are overloaded.
+
+These intrinsics represent histogram-like operations; that is, updating values
+in memory that may not be contiguous, and where multiple elements within a
+single vector may be updating the same value in memory.
+
+The update operation must be specified as part of the intrinsic name. For a
+simple histogram like the following the ``add`` operation would be used.
+
+.. code-block:: c
+
+    void simple_histogram(int *restrict buckets, unsigned *indices, int N, int inc) {
+      for (int i = 0; i < N; ++i)
+        buckets[indices[i]] += inc;
+    }
+
+More update operation types may be added in the future.
+
+::
+
+    declare <8 x i32> @llvm.experimental.vector.histogram.add.v8p0.i32(<8 x ptr> %ptrs, i32 %inc, <8 x i1> %mask)
+    declare <vscale x 2 x i64> @llvm.experimental.vector.histogram.add.nxv2p0.i64(<vscale x 2 x ptr> %ptrs, i64 %inc, <vscale x 2 x i1> %mask)
+
+Arguments:
+""""""""""
+
+The first argument is a vector of pointers to the memory locations to be
+updated. The second argument is a scalar used to update the value from
+memory; it must match the type of value to be updated. The final argument
+is a mask value to exclude locations from being modified.
+
+Semantics:
+""""""""""
+
+The '``llvm.experimental.vector.histogram.*``' intrinsics are used to perform
+updates on potentially overlapping values in memory. The intrinsics represent
+the follow sequence of operations:
+
+1. Gather load from the ``ptrs`` operand, with element type matching that of
+   the ``inc`` operand.
+2. Update of the values loaded from memory. In the case of the ``add``
+   update operation, this means:
+
+   1. Perform a cross-vector histogram operation on the ``ptrs`` operand.
+   2. Multiply the result by the ``inc`` operand.
+   3. Add the result to the values loaded from memory
+3. Scatter the result of the update operation to the memory locations from
+   the ``ptrs`` operand.
+
+The ``mask`` operand will apply to at least the gather and scatter operations.
+
 Matrix Intrinsics
 -----------------
 

llvm/include/llvm/Analysis/TargetTransformInfo.h

@@ -797,6 +797,9 @@ class TargetTransformInfo {
   /// Return true if the target supports strided load.
   bool isLegalStridedLoadStore(Type *DataType, Align Alignment) const;
 
+  // Return true if the target supports masked vector histograms.
+  bool isLegalMaskedVectorHistogram(Type *AddrType, Type *DataType) const;
+
   /// Return true if this is an alternating opcode pattern that can be lowered
   /// to a single instruction on the target. In X86 this is for the addsub
   /// instruction which corrsponds to a Shuffle + Fadd + FSub pattern in IR.
@@ -1883,6 +1886,7 @@ class TargetTransformInfo::Concept {
   virtual bool isLegalMaskedCompressStore(Type *DataType, Align Alignment) = 0;
   virtual bool isLegalMaskedExpandLoad(Type *DataType, Align Alignment) = 0;
   virtual bool isLegalStridedLoadStore(Type *DataType, Align Alignment) = 0;
+  virtual bool isLegalMaskedVectorHistogram(Type *AddrType, Type *DataType) = 0;
   virtual bool isLegalAltInstr(VectorType *VecTy, unsigned Opcode0,
                                unsigned Opcode1,
                                const SmallBitVector &OpcodeMask) const = 0;
@@ -2386,6 +2390,9 @@ class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
   bool isLegalStridedLoadStore(Type *DataType, Align Alignment) override {
     return Impl.isLegalStridedLoadStore(DataType, Alignment);
   }
+  bool isLegalMaskedVectorHistogram(Type *AddrType, Type *DataType) override {
+    return Impl.isLegalMaskedVectorHistogram(AddrType, DataType);
+  }
   bool isLegalAltInstr(VectorType *VecTy, unsigned Opcode0, unsigned Opcode1,
                        const SmallBitVector &OpcodeMask) const override {
     return Impl.isLegalAltInstr(VecTy, Opcode0, Opcode1, OpcodeMask);

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

@@ -316,6 +316,10 @@ class TargetTransformInfoImplBase {
     return false;
   }
 
+  bool isLegalMaskedVectorHistogram(Type *AddrType, Type *DataType) const {
+    return false;
+  }
+
   bool enableOrderedReductions() const { return false; }
 
   bool hasDivRemOp(Type *DataType, bool IsSigned) const { return false; }

llvm/include/llvm/CodeGen/ISDOpcodes.h

@@ -1402,6 +1402,11 @@ enum NodeType {
   // which is later translated to an implicit use in the MIR.
   CONVERGENCECTRL_GLUE,
 
+  // Experimental vector histogram intrinsic
+  // Operands: Input Chain, Inc, Mask, Base, Index, Scale, ID
+  // Output: Output Chain
+  EXPERIMENTAL_VECTOR_HISTOGRAM,
+
   /// BUILTIN_OP_END - This must be the last enum value in this list.
   /// The target-specific pre-isel opcode values start here.
   BUILTIN_OP_END

llvm/include/llvm/CodeGen/SelectionDAG.h

@@ -1526,6 +1526,9 @@ class SelectionDAG {
                            ArrayRef<SDValue> Ops, MachineMemOperand *MMO,
                            ISD::MemIndexType IndexType,
                            bool IsTruncating = false);
+  SDValue getMaskedHistogram(SDVTList VTs, EVT MemVT, const SDLoc &dl,
+                             ArrayRef<SDValue> Ops, MachineMemOperand *MMO,
+                             ISD::MemIndexType IndexType);
 
   SDValue getGetFPEnv(SDValue Chain, const SDLoc &dl, SDValue Ptr, EVT MemVT,
                       MachineMemOperand *MMO);

llvm/include/llvm/CodeGen/SelectionDAGNodes.h

@@ -542,6 +542,7 @@ BEGIN_TWO_BYTE_PACK()
     friend class MaskedLoadStoreSDNode;
     friend class MaskedGatherScatterSDNode;
     friend class VPGatherScatterSDNode;
+    friend class MaskedHistogramSDNode;
 
     uint16_t : NumMemSDNodeBits;
 
@@ -552,6 +553,7 @@ BEGIN_TWO_BYTE_PACK()
     //   MaskedLoadStoreBaseSDNode => enum ISD::MemIndexedMode
     //   VPGatherScatterSDNode => enum ISD::MemIndexType
     //   MaskedGatherScatterSDNode => enum ISD::MemIndexType
+    //   MaskedHistogramSDNode => enum ISD::MemIndexType
     uint16_t AddressingMode : 3;
   };
   enum { NumLSBaseSDNodeBits = NumMemSDNodeBits + 3 };
@@ -564,6 +566,7 @@ BEGIN_TWO_BYTE_PACK()
     friend class MaskedLoadSDNode;
     friend class MaskedGatherSDNode;
     friend class VPGatherSDNode;
+    friend class MaskedHistogramSDNode;
 
     uint16_t : NumLSBaseSDNodeBits;
 
@@ -1420,6 +1423,7 @@ class MemSDNode : public SDNode {
       return getOperand(2);
     case ISD::MGATHER:
     case ISD::MSCATTER:
+    case ISD::EXPERIMENTAL_VECTOR_HISTOGRAM:
       return getOperand(3);
     default:
       return getOperand(1);
@@ -1468,6 +1472,7 @@ class MemSDNode : public SDNode {
     case ISD::EXPERIMENTAL_VP_STRIDED_STORE:
     case ISD::GET_FPENV_MEM:
     case ISD::SET_FPENV_MEM:
+    case ISD::EXPERIMENTAL_VECTOR_HISTOGRAM:
       return true;
     default:
       return N->isMemIntrinsic() || N->isTargetMemoryOpcode();
@@ -2953,6 +2958,34 @@ class MaskedScatterSDNode : public MaskedGatherScatterSDNode {
   }
 };
 
+class MaskedHistogramSDNode : public MemSDNode {
+public:
+  friend class SelectionDAG;
+
+  MaskedHistogramSDNode(unsigned Order, const DebugLoc &DL, SDVTList VTs,
+                        EVT MemVT, MachineMemOperand *MMO,
+                        ISD::MemIndexType IndexType)
+      : MemSDNode(ISD::EXPERIMENTAL_VECTOR_HISTOGRAM, Order, DL, VTs, MemVT,
+                  MMO) {
+    LSBaseSDNodeBits.AddressingMode = IndexType;
+  }
+
+  ISD::MemIndexType getIndexType() const {
+    return static_cast<ISD::MemIndexType>(LSBaseSDNodeBits.AddressingMode);
+  }
+
+  const SDValue &getBasePtr() const { return getOperand(3); }
+  const SDValue &getIndex() const { return getOperand(4); }
+  const SDValue &getMask() const { return getOperand(2); }
+  const SDValue &getScale() const { return getOperand(5); }
+  const SDValue &getInc() const { return getOperand(1); }
+  const SDValue &getIntID() const { return getOperand(6); }
+
+  static bool classof(const SDNode *N) {
+    return N->getOpcode() == ISD::EXPERIMENTAL_VECTOR_HISTOGRAM;
+  }
+};
+
 class FPStateAccessSDNode : public MemSDNode {
 public:
   friend class SelectionDAG;

llvm/include/llvm/IR/Intrinsics.td

@@ -1856,6 +1856,13 @@ def int_experimental_vp_strided_load  : DefaultAttrsIntrinsic<[llvm_anyvector_ty
                                llvm_i32_ty],
                              [ NoCapture<ArgIndex<0>>, IntrNoSync, IntrReadMem, IntrWillReturn, IntrArgMemOnly ]>;
 
+// Experimental histogram
+def int_experimental_vector_histogram_add : DefaultAttrsIntrinsic<[],
+                             [ llvm_anyvector_ty, // Vector of pointers
+                               llvm_anyint_ty,    // Increment
+                               LLVMScalarOrSameVectorWidth<0, llvm_i1_ty>], // Mask
+                             [ IntrArgMemOnly ]>;
+
 // Operators
 let IntrProperties = [IntrNoMem, IntrNoSync, IntrWillReturn] in {
   // Integer arithmetic

llvm/lib/Analysis/TargetTransformInfo.cpp

@@ -513,6 +513,11 @@ bool TargetTransformInfo::isLegalStridedLoadStore(Type *DataType,
   return TTIImpl->isLegalStridedLoadStore(DataType, Alignment);
 }
 
+bool TargetTransformInfo::isLegalMaskedVectorHistogram(Type *AddrType,
+                                                       Type *DataType) const {
+  return TTIImpl->isLegalMaskedVectorHistogram(AddrType, DataType);
+}
+
 bool TargetTransformInfo::enableOrderedReductions() const {
   return TTIImpl->enableOrderedReductions();
 }

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

@@ -9633,6 +9633,44 @@ SDValue SelectionDAG::getMaskedScatter(SDVTList VTs, EVT MemVT, const SDLoc &dl,
   return V;
 }
 
+SDValue SelectionDAG::getMaskedHistogram(SDVTList VTs, EVT MemVT,
+                                         const SDLoc &dl, ArrayRef<SDValue> Ops,
+                                         MachineMemOperand *MMO,
+                                         ISD::MemIndexType IndexType) {
+  assert(Ops.size() == 7 && "Incompatible number of operands");
+
+  FoldingSetNodeID ID;
+  AddNodeIDNode(ID, ISD::EXPERIMENTAL_VECTOR_HISTOGRAM, VTs, Ops);
+  ID.AddInteger(MemVT.getRawBits());
+  ID.AddInteger(getSyntheticNodeSubclassData<MaskedHistogramSDNode>(
+      dl.getIROrder(), VTs, MemVT, MMO, IndexType));
+  ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
+  ID.AddInteger(MMO->getFlags());
+  void *IP = nullptr;
+  if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
+    cast<MaskedGatherSDNode>(E)->refineAlignment(MMO);
+    return SDValue(E, 0);
+  }
+
+  auto *N = newSDNode<MaskedHistogramSDNode>(dl.getIROrder(), dl.getDebugLoc(),
+                                             VTs, MemVT, MMO, IndexType);
+  createOperands(N, Ops);
+
+  assert(N->getMask().getValueType().getVectorElementCount() ==
+             N->getIndex().getValueType().getVectorElementCount() &&
+         "Vector width mismatch between mask and data");
+  assert(isa<ConstantSDNode>(N->getScale()) &&
+         N->getScale()->getAsAPIntVal().isPowerOf2() &&
+         "Scale should be a constant power of 2");
+  assert(N->getInc().getValueType().isInteger() && "Non integer update value");
+
+  CSEMap.InsertNode(N, IP);
+  InsertNode(N);
+  SDValue V(N, 0);
+  NewSDValueDbgMsg(V, "Creating new node: ", this);
+  return V;
+}
+
 SDValue SelectionDAG::getGetFPEnv(SDValue Chain, const SDLoc &dl, SDValue Ptr,
                                   EVT MemVT, MachineMemOperand *MMO) {
   assert(Chain.getValueType() == MVT::Other && "Invalid chain type");

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

@@ -6281,6 +6281,64 @@ void SelectionDAGBuilder::visitConvergenceControl(const CallInst &I,
   }
 }
 
+void SelectionDAGBuilder::visitVectorHistogram(const CallInst &I,
+                                               unsigned IntrinsicID) {
+  // For now, we're only lowering an 'add' histogram.
+  // We can add others later, e.g. saturating adds, min/max.
+  assert(IntrinsicID == Intrinsic::experimental_vector_histogram_add &&
+         "Tried to lower unsupported histogram type");
+  SDLoc sdl = getCurSDLoc();
+  Value *Ptr = I.getOperand(0);
+  SDValue Inc = getValue(I.getOperand(1));
+  SDValue Mask = getValue(I.getOperand(2));
+
+  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  DataLayout TargetDL = DAG.getDataLayout();
+  EVT VT = Inc.getValueType();
+  Align Alignment = DAG.getEVTAlign(VT);
+
+  const MDNode *Ranges = getRangeMetadata(I);
+
+  SDValue Root = DAG.getRoot();
+  SDValue Base;
+  SDValue Index;
+  ISD::MemIndexType IndexType;
+  SDValue Scale;
+  bool UniformBase = getUniformBase(Ptr, Base, Index, IndexType, Scale, this,
+                                    I.getParent(), VT.getScalarStoreSize());
+
+  unsigned AS = Ptr->getType()->getScalarType()->getPointerAddressSpace();
+
+  MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
+      MachinePointerInfo(AS),
+      MachineMemOperand::MOLoad | MachineMemOperand::MOStore,
+      MemoryLocation::UnknownSize, Alignment, I.getAAMetadata(), Ranges);
+
+  if (!UniformBase) {
+    Base = DAG.getConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout()));
+    Index = getValue(Ptr);
+    IndexType = ISD::SIGNED_SCALED;
+    Scale =
+        DAG.getTargetConstant(1, sdl, TLI.getPointerTy(DAG.getDataLayout()));
+  }
+
+  EVT IdxVT = Index.getValueType();
+  EVT EltTy = IdxVT.getVectorElementType();
+  if (TLI.shouldExtendGSIndex(IdxVT, EltTy)) {
+    EVT NewIdxVT = IdxVT.changeVectorElementType(EltTy);
+    Index = DAG.getNode(ISD::SIGN_EXTEND, sdl, NewIdxVT, Index);
+  }
+
+  SDValue ID = DAG.getTargetConstant(IntrinsicID, sdl, MVT::i32);
+
+  SDValue Ops[] = {Root, Inc, Mask, Base, Index, Scale, ID};
+  SDValue Histogram = DAG.getMaskedHistogram(DAG.getVTList(MVT::Other), VT, sdl,
+                                             Ops, MMO, IndexType);
+
+  setValue(&I, Histogram);
+  DAG.setRoot(Histogram);
+}
+
 /// Lower the call to the specified intrinsic function.
 void SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I,
                                              unsigned Intrinsic) {
@@ -7948,6 +8006,11 @@ void SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I,
   case Intrinsic::experimental_convergence_entry:
   case Intrinsic::experimental_convergence_loop:
     visitConvergenceControl(I, Intrinsic);
+    return;
+  case Intrinsic::experimental_vector_histogram_add: {
+    visitVectorHistogram(I, Intrinsic);
+    return;
+  }
   }
 }