[mlir] Fold trivial subtensor / subtensor_insert ops.

[nicolasvasilache]

Summary

Static subtensor / subtensor_insert of the same size as the source / destination tensor and root @[0..0] with strides [1..1] are folded away.

[nicolasvasilache]

Rebase & fix.

mlir/include/mlir/Dialect/StandardOps/IR/Ops.h

 /// comparison predicates.
 bool applyCmpPredicate(CmpIPredicate predicate, const APInt &lhs,
                        const APInt &rhs);
 
 /// Compute `lhs` `pred` `rhs`, where `pred` is one of the known floating point
 /// comparison predicates.
 bool applyCmpPredicate(CmpFPredicate predicate, const APFloat &lhs,
                        const APFloat &rhs);
+
+/// Return true if ofr1 and ofr2 are the same integer constant attribute values
+/// or the same SSA value.
+/// Ignore integer bitwitdh and type mismatch that come from the fact there is
+/// no IndexAttr and that IndexType have no bitwidth.
+bool isEqualConstantIntOrValue(OpFoldResult ofr1, OpFoldResult ofr2);
 } // end namespace mlir
 
 #endif // MLIR_DIALECT_IR_STANDARDOPS_IR_OPS_H

mlir/include/mlir/Dialect/StandardOps/IR/Ops.td

let extraClassDeclaration = extraBaseClassDeclaration # [{
     }
 
     /// Return the number of leading operands before the `offsets`, `sizes` and
     /// and `strides` operands.
     static unsigned getOffsetSizeAndStrideStartOperandIndex() { return 1; }
   }];
 
   let hasCanonicalizer = 1;
+  let hasFolder = 1;
 }
 
 //===----------------------------------------------------------------------===//
 // SubTensorInsertOp
 //===----------------------------------------------------------------------===//
 
 def SubTensorInsertOp : BaseOpWithOffsetSizesAndStrides<
     "subtensor_insert",
std::array<unsigned, 3> getArrayAttrMaxRanks() {
       unsigned rank = getType().getRank();
       return {rank, rank, rank};
     }
 
     /// Return the number of leading operands before the `offsets`, `sizes` and
     /// and `strides` operands.
     static unsigned getOffsetSizeAndStrideStartOperandIndex() { return 2; }
   }];
+  let hasFolder = 1;
 }
 
 //===----------------------------------------------------------------------===//
 // TensorLoadOp
 //===----------------------------------------------------------------------===//
 
 def TensorLoadOp : Std_Op<"tensor_load",
     [SameOperandsAndResultShape, SameOperandsAndResultElementType,

mlir/lib/Dialect/StandardOps/IR/Ops.cpp

 static void dispatchIndexOpFoldResults(ArrayRef<OpFoldResult> ofrs,
                                        SmallVectorImpl<Value> &dynamicVec,
                                        SmallVectorImpl<int64_t> &staticVec,
                                        int64_t sentinel) {
   for (auto ofr : ofrs)
     dispatchIndexOpFoldResult(ofr, dynamicVec, staticVec, sentinel);
 }
 
+/// Return true if ofr1 and ofr2 are the same integer constant attribute values
+/// or the same SSA value.
+/// Ignore integer bitwitdh and type mismatch that come from the fact there is
+/// no IndexAttr and that IndexType have no bitwidth.
+bool mlir::isEqualConstantIntOrValue(OpFoldResult op1, OpFoldResult op2) {
+  auto getConstantIntValue = [](OpFoldResult ofr) -> llvm::Optional<int64_t> {
+    Attribute attr = ofr.dyn_cast<Attribute>();
+    // Note: isa+cast-like pattern allows writing the condition below as 1 line.
+    if (!attr && ofr.get<Value>().getDefiningOp<ConstantOp>())
+      attr = ofr.get<Value>().getDefiningOp<ConstantOp>().getValue();
+    if (auto intAttr = attr.dyn_cast_or_null<IntegerAttr>())
+      return intAttr.getValue().getSExtValue();
+    return llvm::None;
+  };
+  auto cst1 = getConstantIntValue(op1), cst2 = getConstantIntValue(op2);
+  if (cst1 && cst2 && *cst1 == *cst2)
+    return true;
+  auto v1 = op1.dyn_cast<Value>(), v2 = op2.dyn_cast<Value>();
+  return v1 && v2 && v1 == v2;
+}
+
 //===----------------------------------------------------------------------===//
 // StandardOpsDialect Interfaces
 //===----------------------------------------------------------------------===//
 namespace {
 /// This class defines the interface for handling inlining with standard
 /// operations.
 struct StdInlinerInterface : public DialectInlinerInterface {
   using DialectInlinerInterface::DialectInlinerInterface;

 
 void SubTensorOp::getCanonicalizationPatterns(OwningRewritePatternList &results,
                                               MLIRContext *context) {
   results
       .insert<OpWithOffsetSizesAndStridesConstantArgumentFolder<SubTensorOp>>(
           context);
 }
 
+//
+static LogicalResult foldIdentityOffsetSizeAndStrideOpInterface(
+    OffsetSizeAndStrideOpInterface insertOp, ShapedType shapedType) {
+  OpBuilder b(insertOp.getContext());
+  for (OpFoldResult ofr : insertOp.getMixedOffsets())
+    if (!isEqualConstantIntOrValue(ofr, b.getIndexAttr(0)))
+      return failure();
+  // Rank-reducing noops only need to inspect the leading dimensions: llvm::zip
+  // is appropriate.
+  auto shape = shapedType.getShape();
+  for (auto it : llvm::zip(insertOp.getMixedSizes(), shape))
+    if (!isEqualConstantIntOrValue(std::get<0>(it),
+                                   b.getIndexAttr(std::get<1>(it))))
+      return failure();
+  for (OpFoldResult ofr : insertOp.getMixedStrides())
+    if (!isEqualConstantIntOrValue(ofr, b.getIndexAttr(1)))
+      return failure();
+  return success();
+}
+
+OpFoldResult SubTensorOp::fold(ArrayRef<Attribute>) {
+  if (succeeded(foldIdentityOffsetSizeAndStrideOpInterface(*this, getType())))
+    return this->source();
+  return OpFoldResult();
+}
+
 //===----------------------------------------------------------------------===//
 // SubTensorInsertOp
 //===----------------------------------------------------------------------===//
 
 // Build a SubTensorInsertOp with mixed static and dynamic entries.
 void mlir::SubTensorInsertOp::build(OpBuilder &b, OperationState &result,
                                     Value source, Value dest,
                                     ArrayRef<OpFoldResult> offsets,
SmallVector<OpFoldResult> offsetValues = llvm::to_vector<4>(
       llvm::map_range(offsets, [](Value v) -> OpFoldResult { return v; }));
   SmallVector<OpFoldResult> sizeValues = llvm::to_vector<4>(
       llvm::map_range(sizes, [](Value v) -> OpFoldResult { return v; }));
   SmallVector<OpFoldResult> strideValues = llvm::to_vector<4>(
       llvm::map_range(strides, [](Value v) -> OpFoldResult { return v; }));
   build(b, result, source, dest, offsetValues, sizeValues, strideValues);
 }
 
+OpFoldResult SubTensorInsertOp::fold(ArrayRef<Attribute>) {
+  if (succeeded(foldIdentityOffsetSizeAndStrideOpInterface(*this, getType())))
+    return this->source();
+  return OpFoldResult();
+}
+
 //===----------------------------------------------------------------------===//
 // TensorLoadOp
 //===----------------------------------------------------------------------===//
 
 OpFoldResult TensorLoadOp::fold(ArrayRef<Attribute>) {
   if (auto tensorToMemref = memref().getDefiningOp<TensorToMemrefOp>())
     return tensorToMemref.tensor();
   return {};

mlir/test/Dialect/Standard/canonicalize.mlir

 func @subview_of_memcast(%arg : memref<4x6x16x32xi8>) ->
   memref<16x32xi8, affine_map<(d0, d1)[s0] -> (d0 * 32 + d1 + s0)>>{
   %0 = memref_cast %arg : memref<4x6x16x32xi8> to memref<?x?x16x32xi8>
   %1 = subview %0[0, 1, 0, 0] [1, 1, 16, 32] [1, 1, 1, 1] :
     memref<?x?x16x32xi8> to
     memref<16x32xi8, affine_map<(d0, d1)[s0] -> (d0 * 32 + d1 + s0)>>
   return %1 : memref<16x32xi8, affine_map<(d0, d1)[s0] -> (d0 * 32 + d1 + s0)>>
 }
+
+// CHECK-LABEL: func @trivial_subtensor
+//  CHECK-SAME:   %[[ARG0:.[a-z0-9A-Z_]+]]: tensor<4x6x16x32xi8>
+//   CHECK-NOT:   subtensor
+//       CHECK:   return %[[ARG0]] :  tensor<4x6x16x32xi8>
+func @trivial_subtensor(%arg0 : tensor<4x6x16x32xi8>) -> tensor<4x6x16x32xi8> {
+  %0 = subtensor %arg0[0, 0, 0, 0] [4, 6, 16, 32] [1, 1, 1, 1] : tensor<4x6x16x32xi8> to tensor<4x6x16x32xi8>
+  return %0 : tensor<4x6x16x32xi8>
+}
+
+// CHECK-LABEL: func @trivial_subtensor_insert
+//  CHECK-SAME:   %[[ARG0:.[a-z0-9A-Z_]+]]: tensor<4x6x16x32xi8>
+//   CHECK-NOT:   subtensor
+//       CHECK:   return %[[ARG0]] :  tensor<4x6x16x32xi8>
+func @trivial_subtensor_insert(%arg0 : tensor<4x6x16x32xi8>, %arg1 : tensor<4x6x16x32xi8>) -> tensor<4x6x16x32xi8> {
+  %0 = subtensor_insert %arg0 into %arg1[0, 0, 0, 0] [4, 6, 16, 32] [1, 1, 1, 1] : tensor<4x6x16x32xi8> into tensor<4x6x16x32xi8>
+  return %0 : tensor<4x6x16x32xi8>
+}
+