Enables inferring return types for Shape op if possible

[jpienaar]

mlir/lib/Dialect/Shape/IR/Shape.cpp

413

The number of operands may not have been verified here yet, lets return failure if there are not 2 operands

424

Lets check here that it is one of the given types just as haven't verified (lets chat about that later this week too)

Seems like a few of these could be deduped, which would help if we have shared functions here to refactor later.

[Chia-hungDuan]

Address review comment and fix the isCompatibleReturnType of MaxOp and MinOp

[Chia-hungDuan]

Add one more type check for ShapeOfOp

[Chia-hungDuan]

mlir/lib/Dialect/Shape/IR/Shape.cpp

413

Added check for size of operands

[jpienaar]

mlir/include/mlir/Dialect/Shape/IR/ShapeOps.td

31

We don't have a formatgter for this yet, but the convention we've been following is to have : ShapeOp on the line in line with the def

def Shape_AddOp : Shape_Op<"add", [

in some cases the [ and traits on the next, but mostly it opens on the former.

mlir/lib/Dialect/Shape/IR/Shape.cpp

424

May be a case of missing change upload again , I'll double check

426

Mmm, this seems to be the same as

return l.front().isa<SizeType, IndexType>() || r.front().isa<SizeType, IndexType>() ?

1191

Aren't these the only two allowed types for this operand?

[Chia-hungDuan]

Remove the type/size check for operands in inferReturnTypes and addressed the review comments

[Chia-hungDuan]

mlir/lib/Dialect/Shape/IR/Shape.cpp

426

I guess you were saying
return l.front().isa<SizeType, IndexType>() && r.front().isa<SizeType, IndexType>()
Done.

1191

Yes,

let results = (outs Shape_ShapeOrSizeType:$result);

[jpienaar]

mlir/include/mlir/Dialect/Shape/IR/ShapeOps.td

54

Perhaps (here and below), Returns when two result types are compatible for this op; method used by InferTypeOpInterface ?

mlir/lib/Dialect/Shape/IR/Shape.cpp

421

Could we make a static helper function and just call it from the others? (this seems like the same code below)

1181

If we had

lhs.isa<IndexType, SizeType> && rhs.isa<SizeType, IndexType>()

but lhs.isa<IndexType> && rhs.isa<SizeType>, then this would fail, but seems fine/non-error. Or am I missing it?

1263

Could we just use operands[0].getType() here rather than getting one?

[Chia-hungDuan]

Address review comment

[Chia-hungDuan]

mlir/lib/Dialect/Shape/IR/Shape.cpp

1181

let results = (outs Shape_ShapeOrSizeType:$result);

The result of JoinOp is Shape_ShapeOrSizeType. Or do we need to update the semantic of JoinOp to make it also be able to produce IndexType?

1263

It should be isa<ShapeType> here. Fixed

[Chia-hungDuan]

Rebase and minor update commit message

mlir/include/mlir/Dialect/Shape/IR/ShapeOps.td

 //===----------------------------------------------------------------------===//
 // Shape op definitions
 //===----------------------------------------------------------------------===//
 
 // Base class for the operation in this dialect
 class Shape_Op<string mnemonic, list<OpTrait> traits = []> :
     Op<ShapeDialect, mnemonic, traits>;
 
-def Shape_AddOp : Shape_Op<"add", [Commutative, NoSideEffect]> {
+def Shape_AddOp
+    : Shape_Op<"add", [Commutative, NoSideEffect,
+                       DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
   let summary = "Addition of sizes and indices";
   let description = [{
     Adds two sizes or indices. If either operand is an error it will be
     propagated to the result. The operands can be of type `size` or `index`. If
     at least one of the operands can hold an error, i.e. if it is of type `size`,
     the result must be of type `size`. If error propagation is not possible
     because both operands are of type `index` then the result may be of type
     `size` or `index`.
   }];
 
   let arguments = (ins Shape_SizeOrIndexType:$lhs, Shape_SizeOrIndexType:$rhs);
   let results = (outs Shape_SizeOrIndexType:$result);
 
   let assemblyFormat = [{
     $lhs `,` $rhs attr-dict `:` type($lhs) `,` type($rhs) `->` type($result)
   }];
 
   let verifier = [{ return verifySizeOrIndexOp(*this); }];
+
+  let extraClassDeclaration = [{
+    /// InferTypeOpInterface:
+    static bool isCompatibleReturnTypes(TypeRange l, TypeRange r);
+  }];
 }
 
 def Shape_BroadcastOp : Shape_Op<"broadcast", [Commutative, NoSideEffect]> {
   let summary = "Returns the broadcasted output shape of two or more inputs";
   let description = [{
     Returns the broadcasted shape for input shapes or extent tensors. The rest
     of this description is simplified for the 2 input case but can be extended
     to more inputs. Both operands can be of type `shape.shape` or
let description = [{
     the error value) the behavior may be undefined. The optional string
     attribute can be used to describe the error case.
   }];
 
   let arguments = (ins Variadic<Shape_ShapeOrExtentTensorType>:$shapes,
                        OptionalAttr<StrAttr>:$error);
   let results = (outs Shape_ShapeOrExtentTensorType:$result);
 
+  let builders = [OpBuilder<(ins "Value":$shape)>];
+
   let assemblyFormat = [{
     $shapes attr-dict `:` type($shapes) `->` type($result)
   }];
 
   let builders = [OpBuilder<(ins "::mlir::Type":$result,
                                 "::mlir::Value":$lhs, "::mlir::Value":$rhs,
                                 "/*optional*/ ::mlir::StringAttr":$error), [{
       build($_builder, $_state, result, ::llvm::makeArrayRef({lhs, rhs}), error);
def Shape_ConstSizeOp : Shape_Op<"const_size", [
   let results = (outs Shape_SizeType:$result);
 
   let builders = [OpBuilder<(ins "int64_t":$value)>];
 
   let assemblyFormat = "$value attr-dict";
   let hasFolder = 1;
 }
 
-def Shape_DivOp : Shape_Op<"div", [NoSideEffect]> {
+def Shape_DivOp
+    : Shape_Op<"div", [NoSideEffect,
+                       DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
   let summary = "Division of sizes and indices";
   let description = [{
     Divides two sizes or indices. If either operand is an error it will be
     propagated to the result. The operands can be of type `size` or `index`.
     If at least one of the operands can hold an error, i.e. if it is of type
     `size`, the result must be of type `size`. If error propagation is not
     possible because both operands are of type `index` then the result may be
     of type  `size` or `index`. If both operands and result are of type `index`,
def Shape_DivOp
   let results = (outs Shape_SizeOrIndexType:$result);
 
   let assemblyFormat = [{
     $lhs `,` $rhs attr-dict `:` type($lhs) `,` type($rhs) `->` type($result)
   }];
 
   let verifier = [{ return ::verifySizeOrIndexOp(*this); }];
   let hasFolder = 1;
+
+  let extraClassDeclaration = [{
+    /// InferTypeOpInterface:
+    static bool isCompatibleReturnTypes(TypeRange l, TypeRange r);
+  }];
 }
 
-def Shape_ShapeEqOp : Shape_Op<"shape_eq", [NoSideEffect, Commutative,
-                                            InferTypeOpInterface]> {
+def Shape_ShapeEqOp
+    : Shape_Op<"shape_eq", [NoSideEffect, Commutative, InferTypeOpInterface]> {
   let summary = "Returns whether the input shapes or extent tensors are equal";
   let description = [{
     Takes one or more shape or extent tensor operands and determines whether
     they are equal. When extent tensors are compared to shapes they are regarded
     as their equivalent non-error shapes. Error shapes can be tested for
     equality like any other shape value, meaning that the error value is equal
     to itself.
   }];
def Shape_IsBroadcastableOp : Shape_Op<"is_broadcastable",
   }];
 
   let hasFolder = 1;
   let hasCanonicalizer = 1;
 
   let assemblyFormat = "$shapes attr-dict `:` type($shapes)";
 }
 
-def Shape_RankOp : Shape_Op<"rank", [NoSideEffect]> {
+def Shape_RankOp
+    : Shape_Op<"rank", [NoSideEffect,
+                        DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
   let summary = "Gets the rank of a shape";
   let description = [{
     Returns the rank of the shape or extent tensor, i.e. the number of extents.
   }];
 
   let arguments = (ins Shape_ShapeOrExtentTensorType:$shape);
   let results = (outs Shape_SizeOrIndexType:$rank);
 
   let assemblyFormat = "$shape attr-dict `:` type($shape) `->` type($rank)";
 
   let hasFolder = 1;
   let hasCanonicalizer = 1;
   let verifier = [{ return ::verifySizeOrIndexOp(*this); }];
+
+  let extraClassDeclaration = [{
+    /// InferTypeOpInterface:
+    static bool isCompatibleReturnTypes(TypeRange l, TypeRange r);
+  }];
 }
 
 def Shape_ToExtentTensorOp : Shape_Op<"to_extent_tensor", [NoSideEffect]> {
   let summary = "Creates a dimension tensor from a shape";
   let description = [{
     Converts a shape to a 1D integral tensor of extents. The number of elements
     in the tensor equals the rank of the shape, and the elements equal the
     extents of the shape.
 
     If the shape represents an error, this op's behavior is undefined.
   }];
 
   let arguments = (ins Shape_ShapeOrExtentTensorType:$input);
   let results = (outs IndexTensor:$result);
 
   let assemblyFormat = "$input attr-dict `:` type($input) `->` type($result)";
 
   let hasFolder = 1;
 }
 
-def Shape_GetExtentOp : Shape_Op<"get_extent", [NoSideEffect]> {
+def Shape_GetExtentOp
+    : Shape_Op<"get_extent",
+               [NoSideEffect,
+                DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
   let summary = "Gets the specified extent from a shape or extent tensor";
   let description = [{
     Gets the extent indexed by `dim` from the `shape` operand. If the shape is
     an error then it returns an invalid size.
   }];
   let arguments = (ins Shape_ShapeOrExtentTensorType:$shape,
                        Shape_SizeOrIndexType:$dim);
   let results = (outs Shape_SizeOrIndexType:$extent);
   let assemblyFormat = "$shape `,` $dim attr-dict `:` type($shape) `,` type($dim) `->` "
                        "type($extent)";
 
   let builders = [
     // Builder that allows passing a constant dimension as a simple integer.
     OpBuilder<(ins "Value":$shape, "int64_t":$dim)>
   ];
 
   let extraClassDeclaration = [{
     /// Get the `dim` value as integer if it is constant.
     Optional<int64_t> getConstantDim();
+    /// InferTypeOpInterface:
+    static bool isCompatibleReturnTypes(TypeRange l, TypeRange r);
   }];
 
   let hasFolder = 1;
   let verifier = [{ return ::verifySizeOrIndexOp(*this); }];
 }
 
 def Shape_IndexToSizeOp : Shape_Op<"index_to_size", [NoSideEffect]> {
   let summary = "Converts a standard index to a shape size";
def Shape_IndexToSizeOp : Shape_Op<"index_to_size", [NoSideEffect]> {
   let results = (outs Shape_SizeType:$result);
 
   let assemblyFormat = "$arg attr-dict";
 
   let hasFolder = 1;
   let hasCanonicalizer = 1;
 }
 
-def Shape_JoinOp : Shape_Op<"join", [Commutative]> {
+def Shape_JoinOp
+    : Shape_Op<"join", [Commutative,
+                        DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
   let summary = "Returns the least general shape.shape of its operands";
   let description = [{
     An operation that computes the least general shape of input operands.
     This effectively asserts that corresponding static dimensions are equal.
     The behavior is to match each element of the `shape.shape` and propagate the
     most restrictive information, returning an invalid shape if there are
     contradictory requirements. E.g., using pseudo code
 
def Shape_JoinOp
   let arguments = (ins Shape_ShapeOrSizeType:$arg0, Shape_ShapeOrSizeType:$arg1,
                    OptionalAttr<StrAttr>:$error);
   let results = (outs Shape_ShapeOrSizeType:$result);
 
   let assemblyFormat = [{
     $arg0 `,` $arg1 (`,` `error` `=` $error^)? attr-dict `:`
       type($arg0) `,` type($arg1) `->` type($result)
   }];
+
+  let extraClassDeclaration = [{
+    // InferTypeOpInterface:
+    static bool isCompatibleReturnTypes(TypeRange l, TypeRange r);
+  }];
 }
 
-def Shape_MaxOp : Shape_Op<"max", [Commutative, NoSideEffect]> {
+def Shape_MaxOp
+    : Shape_Op<"max", [Commutative, NoSideEffect,
+                       DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
   let summary = "Elementwise maximum";
   let description = [{
     Computes the elementwise maximum of two sizes or shapes with equal ranks.
     If either operand is an error, then an error will be propagated to the
     result. If the input types mismatch or the ranks do not match, then the
     result is an error.
   }];
 
   let arguments = (ins Shape_ShapeOrSizeType:$lhs, Shape_ShapeOrSizeType:$rhs);
   let results = (outs Shape_ShapeOrSizeType:$result);
 
   let assemblyFormat = [{
     $lhs `,` $rhs attr-dict `:` type($lhs) `,` type($rhs) `->` type($result)
   }];
 
   let hasFolder = 1;
+
+  let extraClassDeclaration = [{
+    // InferTypeOpInterface:
+    static bool isCompatibleReturnTypes(TypeRange l, TypeRange r);
+  }];
 }
 
-def Shape_MinOp : Shape_Op<"min", [Commutative, NoSideEffect]> {
+def Shape_MinOp
+    : Shape_Op<"min", [Commutative, NoSideEffect,
+                       DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
   let summary = "Elementwise minimum";
   let description = [{
     Computes the elementwise minimum of two sizes or shapes with equal ranks.
     If either operand is an error, then an error will be propagated to the
     result. If the input types mismatch or the ranks do not match, then the
     result is an error.
   }];
 
   let arguments = (ins Shape_ShapeOrSizeType:$lhs, Shape_ShapeOrSizeType:$rhs);
   let results = (outs Shape_ShapeOrSizeType:$result);
 
   let assemblyFormat = [{
     $lhs `,` $rhs attr-dict `:` type($lhs) `,` type($rhs) `->` type($result)
   }];
 
   let hasFolder = 1;
+
+  let extraClassDeclaration = [{
+    // InferTypeOpInterface:
+    static bool isCompatibleReturnTypes(TypeRange l, TypeRange r);
+  }];
 }
 
-def Shape_MulOp : Shape_Op<"mul", [Commutative, NoSideEffect]> {
+def Shape_MulOp
+    : Shape_Op<"mul", [Commutative, NoSideEffect,
+                       DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
   let summary = "Multiplication of sizes and indices";
   let description = [{
     Multiplies two sizes or indices. If either operand is an error it will be
     propagated to the result. The operands can be of type `size` or `index`. If
     at least one of the operands can hold an error, i.e. if it is of type `size`,
     the result must be of type `size`. If error propagation is not possible
     because both operands are of type `index` then the result may be of type
     `size` or `index`.
   }];
 
   let arguments = (ins Shape_SizeOrIndexType:$lhs, Shape_SizeOrIndexType:$rhs);
   let results = (outs Shape_SizeOrIndexType:$result);
 
   let assemblyFormat = [{
     $lhs `,` $rhs attr-dict `:` type($lhs) `,` type($rhs) `->` type($result)
   }];
 
   let verifier = [{ return ::verifySizeOrIndexOp(*this); }];
   let hasFolder = 1;
+
+  let extraClassDeclaration = [{
+    /// InferTypeOpInterface:
+    static bool isCompatibleReturnTypes(TypeRange l, TypeRange r);
+  }];
 }
 
-def Shape_NumElementsOp : Shape_Op<"num_elements", [NoSideEffect]> {
+def Shape_NumElementsOp
+    : Shape_Op<"num_elements",
+               [NoSideEffect,
+                DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
   let summary = "Returns the number of elements for a given shape";
   let description = [{
     Returns the number of elements for a given shape which is the product of its
     extents. If the argument is of type `shape` then the result will be of type
     `size` and potential errors will be propagated. Otherwise, if the argument
     is and extent tensor `tensor<?xindex>` then the result will be of type
     `index`.
   }];
 
   let arguments = (ins Shape_ShapeOrExtentTensorType:$shape);
   let results = (outs Shape_SizeOrIndexType:$result);
 
-  let builders = [OpBuilder<(ins "Value":$shape)>];
-
   let assemblyFormat = "$shape attr-dict `:` type($shape) `->` type($result)";
 
   let hasFolder = 1;
   let verifier = [{ return ::verifySizeOrIndexOp(*this); }];
+  let extraClassDeclaration = [{
+    /// InferTypeOpInterface:
+    static bool isCompatibleReturnTypes(TypeRange l, TypeRange r);
+  }];
 }
 
 def Shape_ReduceOp : Shape_Op<"reduce",
     [SingleBlockImplicitTerminator<"YieldOp">]> {
   let summary = "Returns an expression reduced over a shape or extent tensor";
   let description = [{
     An operation that takes as input a shape or extent tensor, and a number of
     initial values. This operation has a region that is applied repeatedly for
def Shape_ReduceOp : Shape_Op<"reduce",
 
   let builders = [OpBuilder<(ins "Value":$shape, "ValueRange":$initVals)>];
 
   let verifier = [{ return ::verify(*this); }];
   let printer = [{ return ::print(p, *this); }];
   let parser = [{ return ::parse$cppClass(parser, result); }];
 }
 
-def Shape_ShapeOfOp : Shape_Op<"shape_of", [NoSideEffect]> {
+def Shape_ShapeOfOp
+    : Shape_Op<"shape_of",
+               [NoSideEffect,
+                DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
   let summary = "Returns shape of a value or shaped type operand";
 
   let description = [{
     The operation takes a value or a shaped operand as an argument and it
     returns a shape or extent tensor.
   }];
 
   let arguments = (ins AnyTypeOf<[AnyShaped, Shape_ValueShapeType]>:$arg);
   let results = (outs Shape_ShapeOrExtentTensorType:$result);
 
   let assemblyFormat = "$arg attr-dict `:` type($arg) `->` type($result)";
 
-  let builders = [OpBuilder<(ins "Value":$arg)>];
-
   let verifier = [{ return ::verifyShapeOrExtentTensorOp(*this); }];
   let hasCanonicalizer = 1;
   let hasFolder = 1;
+
+  let extraClassDeclaration = [{
+    // InferTypeOpInterface:
+    static bool isCompatibleReturnTypes(TypeRange l, TypeRange r);
+  }];
 }
 
 def Shape_SizeToIndexOp : Shape_Op<"size_to_index", [NoSideEffect]> {
   let summary = "Casts between index types of the shape and standard dialect";
   let description = [{
     Converts a `shape.size` to a standard index. This operation and its
     inverse, `index_to_size`, facilitate index conversion between the standard
     and the shape dialect. The behavior is undefined for unknown and invalid
let extraClassDeclaration = [{
     };
   }];
 
   let hasCanonicalizer = 1;
   let hasFolder = 1;
   let verifier = [{ return ::verify(*this); }];
 }
 
-def Shape_CstrEqOp : Shape_Op<"cstr_eq", [Commutative, InferTypeOpInterface]> {
+def Shape_CstrEqOp
+    : Shape_Op<"cstr_eq", [Commutative, InferTypeOpInterface]> {
   let summary = "Determines if all input shapes are equal";
   let description = [{
     Given 1 or more input shapes, determine if all shapes are the exact same.
 
     "cstr" operations represent runtime assertions.
 
     Example:
     ```mlir

mlir/lib/Dialect/Shape/IR/Shape.cpp

 
 #include "mlir/Dialect/StandardOps/IR/Ops.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/Dialect/Traits.h"
 #include "mlir/IR/Builders.h"
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/DialectImplementation.h"
 #include "mlir/IR/PatternMatch.h"
+#include "mlir/IR/TypeUtilities.h"
 #include "mlir/Transforms/InliningUtils.h"
 #include "llvm/ADT/SmallString.h"
 #include "llvm/ADT/TypeSwitch.h"
 #include "llvm/Support/raw_ostream.h"
 
 using namespace mlir;
 using namespace mlir::shape;
 
void AssumingOp::build(
 
   SmallVector<Type, 2> assumingTypes;
   for (Value v : yieldValues)
     assumingTypes.push_back(v.getType());
   result.addTypes(assumingTypes);
 }
 
 //===----------------------------------------------------------------------===//
+// AddOp
+//===----------------------------------------------------------------------===//
+
+LogicalResult mlir::shape::AddOp::inferReturnTypes(
+    MLIRContext *context, Optional<Location> location, ValueRange operands,
+    DictionaryAttr attributes, RegionRange regions,
+    SmallVectorImpl<Type> &inferredReturnTypes) {
+  if (operands.size() != 2)
+    return failure();
+  if (operands[0].getType().isa<SizeType>() ||
+      operands[1].getType().isa<SizeType>())
+    inferredReturnTypes.assign({SizeType::get(context)});
+  else
+    inferredReturnTypes.assign({IndexType::get(context)});
+  return success();
+}
+
+bool mlir::shape::AddOp::isCompatibleReturnTypes(TypeRange l, TypeRange r) {
+  if (l.size() != 1 || r.size() != 1)
+    return false;
+  if (!l.front().isa<SizeType, IndexType>() ||
+      !r.front().isa<SizeType, IndexType>())
+    return false;
+
+  // SizeType is compatible with IndexType.
+  return true;
+}
+
+//===----------------------------------------------------------------------===//
 // AssumingAllOp
 //===----------------------------------------------------------------------===//
 
 namespace {
 struct AssumingAllToCstrEqCanonicalization
     : public OpRewritePattern<AssumingAllOp> {
   using OpRewritePattern<AssumingAllOp>::OpRewritePattern;
 
OpFoldResult DivOp::fold(ArrayRef<Attribute> operands) {
   if (quotient.isNegative() && !remainder.isNullValue()) {
     quotient -= 1;
   }
 
   Type indexTy = IndexType::get(getContext());
   return IntegerAttr::get(indexTy, quotient);
 }
 
+LogicalResult mlir::shape::DivOp::inferReturnTypes(
+    MLIRContext *context, Optional<Location> location, ValueRange operands,
+    DictionaryAttr attributes, RegionRange regions,
+    SmallVectorImpl<Type> &inferredReturnTypes) {
+  if (operands.size() != 2)
+    return failure();
+  if (operands[0].getType().isa<SizeType>() ||
+      operands[1].getType().isa<SizeType>())
+    inferredReturnTypes.assign({SizeType::get(context)});
+  else
+    inferredReturnTypes.assign({IndexType::get(context)});
+  return success();
+}
+
+bool mlir::shape::DivOp::isCompatibleReturnTypes(TypeRange l, TypeRange r) {
+  if (l.size() != 1 || r.size() != 1)
+    return false;
+  if (!l.front().isa<SizeType, IndexType>() ||
+      !r.front().isa<SizeType, IndexType>())
+    return false;
+
+  // SizeType is compatible with IndexType.
+  return true;
+}
+
 //===----------------------------------------------------------------------===//
 // ShapeEqOp
 //===----------------------------------------------------------------------===//
 
 OpFoldResult ShapeEqOp::fold(ArrayRef<Attribute> operands) {
   bool allSame = true;
   if (!operands.empty() && !operands[0])
     return {};
if (shape.getType().isa<ShapeType>()) {
     build(builder, result, builder.getType<SizeType>(), shape, dim);
   } else {
     Value dim =
         builder.create<ConstantOp>(loc, builder.getIndexType(), dimAttr);
     build(builder, result, builder.getIndexType(), shape, dim);
   }
 }
 
+LogicalResult mlir::shape::GetExtentOp::inferReturnTypes(
+    MLIRContext *context, Optional<Location> location, ValueRange operands,
+    DictionaryAttr attributes, RegionRange regions,
+    SmallVectorImpl<Type> &inferredReturnTypes) {
+  inferredReturnTypes.assign({IndexType::get(context)});
+  return success();
+}
+
+bool mlir::shape::GetExtentOp::isCompatibleReturnTypes(TypeRange l,
+                                                       TypeRange r) {
+  if (l.size() != 1 || r.size() != 1)
+    return false;
+  if (!l.front().isa<SizeType, IndexType>() ||
+      !r.front().isa<SizeType, IndexType>())
+    return false;
+
+  // SizeType is compatible with IndexType.
+  return true;
+}
+
 //===----------------------------------------------------------------------===//
 // IsBroadcastableOp
 //===----------------------------------------------------------------------===//
 
 void IsBroadcastableOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
                                                     MLIRContext *context) {
   patterns.add<RemoveDuplicateOperandsPattern<IsBroadcastableOp>>(context);
 }
 
 OpFoldResult IsBroadcastableOp::fold(ArrayRef<Attribute> operands) {
   // Can always broadcast fewer than two shapes.
   if (operands.size() < 2) {
     return BoolAttr::get(getContext(), true);
   }
 
   return nullptr;
 }
 
 //===----------------------------------------------------------------------===//
+// JoinOp
+//===----------------------------------------------------------------------===//
+
+LogicalResult mlir::shape::JoinOp::inferReturnTypes(
+    MLIRContext *context, Optional<Location> location, ValueRange operands,
+    DictionaryAttr attributes, RegionRange regions,
+    SmallVectorImpl<Type> &inferredReturnTypes) {
+  if (operands.size() != 2)
+    return failure();
+
+  // The type of operands must be consistent.
+  if (operands[0].getType() != operands[1].getType())
+    return failure();
+
+  if (operands[0].getType().isa<ShapeType>())
+    inferredReturnTypes.assign({ShapeType::get(context)});
+  else
+    inferredReturnTypes.assign({SizeType::get(context)});
+  return success();
+}
+
+bool mlir::shape::JoinOp::isCompatibleReturnTypes(TypeRange l, TypeRange r) {
+  if (l.size() != 1 || r.size() != 1)
+    return false;
+  if (l == r)
+    return true;
+
+  Type lhs = l.front();
+  Type rhs = r.front();
+
+  if (lhs != rhs)
+    return false;
+
+  if (lhs.isa<SizeType>() || lhs.isa<ShapeType>())
+    return true;
+
+  if (succeeded(verifyCompatibleShapes({lhs, rhs})))
+    return true;
+  return false;
+}
+
+//===----------------------------------------------------------------------===//
 // RankOp
 //===----------------------------------------------------------------------===//
 
 OpFoldResult shape::RankOp::fold(ArrayRef<Attribute> operands) {
   auto shape = operands[0].dyn_cast_or_null<DenseIntElementsAttr>();
   if (!shape)
     return {};
   int64_t rank = shape.getNumElements();

 };
 } // namespace
 
 void shape::RankOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
                                                 MLIRContext *context) {
   patterns.add<RankShapeOfCanonicalizationPattern>(context);
 }
 
+LogicalResult mlir::shape::RankOp::inferReturnTypes(
+    MLIRContext *context, Optional<Location> location, ValueRange operands,
+    DictionaryAttr attributes, RegionRange regions,
+    SmallVectorImpl<Type> &inferredReturnTypes) {
+  if (operands.size() != 1)
+    return failure();
+  if (operands[0].getType().isa<SizeType>())
+    inferredReturnTypes.assign({SizeType::get(context)});
+  else
+    inferredReturnTypes.assign({IndexType::get(context)});
+  return success();
+}
+
+bool mlir::shape::RankOp::isCompatibleReturnTypes(TypeRange l, TypeRange r) {
+  if (l.size() != 1 || r.size() != 1)
+    return false;
+  if (!l.front().isa<SizeType, IndexType>() ||
+      !r.front().isa<SizeType, IndexType>())
+    return false;
+
+  // SizeType is compatible with IndexType.
+  return true;
+}
+
 //===----------------------------------------------------------------------===//
 // NumElementsOp
 //===----------------------------------------------------------------------===//
 
 OpFoldResult NumElementsOp::fold(ArrayRef<Attribute> operands) {
 
   // Fold only when argument constant.
   Attribute shape = operands[0];
   if (!shape)
     return {};
 
   APInt product(64, 1);
   for (auto value : shape.cast<DenseIntElementsAttr>())
     product *= value;
   Builder builder(getContext());
   return builder.getIndexAttr(product.getLimitedValue());
 }
 
-void NumElementsOp::build(OpBuilder &builder, OperationState &result,
-                          Value shape) {
-  if (shape.getType().isa<ShapedType>()) {
-    auto type = builder.getIndexType();
-    return build(builder, result, type, shape);
-  }
-  auto type = SizeType::get(builder.getContext());
-  return build(builder, result, type, shape);
+LogicalResult mlir::shape::NumElementsOp::inferReturnTypes(
+    MLIRContext *context, Optional<Location> location, ValueRange operands,
+    DictionaryAttr attributes, RegionRange regions,
+    SmallVectorImpl<Type> &inferredReturnTypes) {
+  if (operands.size() != 1)
+    return failure();
+  if (operands[0].getType().isa<ShapeType>())
+    inferredReturnTypes.assign({SizeType::get(context)});
+  else
+    inferredReturnTypes.assign({IndexType::get(context)});
+  return success();
+
+bool mlir::shape::NumElementsOp::isCompatibleReturnTypes(TypeRange l,
+                                                         TypeRange r) {
+  if (l.size() != 1 || r.size() != 1)
+    return false;
+  if (!l.front().isa<SizeType, IndexType>() ||
+      !r.front().isa<SizeType, IndexType>())
+    return false;
+
+  // SizeType is compatible with IndexType.
+  return true;
 }
 
 //===----------------------------------------------------------------------===//
 // MaxOp
 //===----------------------------------------------------------------------===//
 
 OpFoldResult MaxOp::fold(llvm::ArrayRef<mlir::Attribute> operands) {
   // If operands are equal, just propagate one.
   if (lhs() == rhs())
     return lhs();
   return nullptr;
 }
 
+LogicalResult mlir::shape::MaxOp::inferReturnTypes(
+    MLIRContext *context, Optional<Location> location, ValueRange operands,
+    DictionaryAttr attributes, RegionRange regions,
+    SmallVectorImpl<Type> &inferredReturnTypes) {
+  if (operands.size() != 2)
+    return failure();
+  if (operands[0].getType() == operands[1].getType())
+    inferredReturnTypes.assign({operands[0].getType()});
+  else
+    inferredReturnTypes.assign({SizeType::get(context)});
+  return success();
+}
+
+bool mlir::shape::MaxOp::isCompatibleReturnTypes(TypeRange l, TypeRange r) {
+  if (l.size() != 1 || r.size() != 1)
+    return false;
+  if (l.front().isa<ShapeType>() && r.front().isa<ShapeType>())
+    return true;
+  if (l.front().isa<SizeType>() && r.front().isa<SizeType>())
+    return true;
+  return false;
+}
+
 //===----------------------------------------------------------------------===//
 // MinOp
 //===----------------------------------------------------------------------===//
 
 OpFoldResult MinOp::fold(llvm::ArrayRef<mlir::Attribute> operands) {
   // If operands are equal, just propagate one.
   if (lhs() == rhs())
     return lhs();
   return nullptr;
 }
 
+LogicalResult mlir::shape::MinOp::inferReturnTypes(
+    MLIRContext *context, Optional<Location> location, ValueRange operands,
+    DictionaryAttr attributes, RegionRange regions,
+    SmallVectorImpl<Type> &inferredReturnTypes) {
+  if (operands.size() != 2)
+    return failure();
+  if (operands[0].getType() == operands[1].getType())
+    inferredReturnTypes.assign({operands[0].getType()});
+  else
+    inferredReturnTypes.assign({SizeType::get(context)});
+  return success();
+}
+
+bool mlir::shape::MinOp::isCompatibleReturnTypes(TypeRange l, TypeRange r) {
+  if (l.size() != 1 || r.size() != 1)
+    return false;
+  if (l.front().isa<ShapeType>() && r.front().isa<ShapeType>())
+    return true;
+  if (l.front().isa<SizeType>() && r.front().isa<SizeType>())
+    return true;
+  return false;
+}
+
 //===----------------------------------------------------------------------===//
 // MulOp
 //===----------------------------------------------------------------------===//
 
 OpFoldResult MulOp::fold(ArrayRef<Attribute> operands) {
   auto lhs = operands[0].dyn_cast_or_null<IntegerAttr>();
   if (!lhs)
     return nullptr;
   auto rhs = operands[1].dyn_cast_or_null<IntegerAttr>();
   if (!rhs)
     return nullptr;
   APInt folded = lhs.getValue() * rhs.getValue();
   Type indexTy = IndexType::get(getContext());
   return IntegerAttr::get(indexTy, folded);
 }
 
+LogicalResult mlir::shape::MulOp::inferReturnTypes(
+    MLIRContext *context, Optional<Location> location, ValueRange operands,
+    DictionaryAttr attributes, RegionRange regions,
+    SmallVectorImpl<Type> &inferredReturnTypes) {
+  if (operands.size() != 2)
+    return failure();
+  if (operands[0].getType().isa<SizeType>() ||
+      operands[1].getType().isa<SizeType>())
+    inferredReturnTypes.assign({SizeType::get(context)});
+  else
+    inferredReturnTypes.assign({IndexType::get(context)});
+  return success();
+}
+
+bool mlir::shape::MulOp::isCompatibleReturnTypes(TypeRange l, TypeRange r) {
+  if (l.size() != 1 || r.size() != 1)
+    return false;
+  if (!l.front().isa<SizeType, IndexType>() ||
+      !r.front().isa<SizeType, IndexType>())
+    return false;
+
+  // SizeType is compatible with IndexType.
+  return true;
+}
 //===----------------------------------------------------------------------===//
 // ShapeOfOp
 //===----------------------------------------------------------------------===//
 
 OpFoldResult ShapeOfOp::fold(ArrayRef<Attribute>) {
   auto type = getOperand().getType().dyn_cast<ShapedType>();
   if (!type || !type.hasStaticShape())
     return nullptr;
   Builder builder(getContext());
   return builder.getIndexTensorAttr(type.getShape());
 }
 
-void ShapeOfOp::build(OpBuilder &builder, OperationState &result, Value arg) {
-  if (auto shapedTy = arg.getType().dyn_cast<ShapedType>()) {
-    int64_t rank =
-        shapedTy.hasRank() ? shapedTy.getRank() : ShapedType::kDynamicSize;
-    Type indexTy = builder.getIndexType();
-    Type extentTensorTy = RankedTensorType::get({rank}, indexTy);
-    return ShapeOfOp::build(builder, result, extentTensorTy, arg);
-  }
-  Type shapeTy = builder.getType<ShapeType>();
-  return ShapeOfOp::build(builder, result, shapeTy, arg);
-}
-
 namespace {
 struct ShapeOfWithTensor : public OpRewritePattern<shape::ShapeOfOp> {
   using OpRewritePattern<shape::ShapeOfOp>::OpRewritePattern;
 
   LogicalResult matchAndRewrite(shape::ShapeOfOp op,
                                 PatternRewriter &rewriter) const override {
     if (!op.arg().getType().isa<ShapedType>())
       return failure();

 };
 } // namespace
 
 void ShapeOfOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
                                             MLIRContext *context) {
   patterns.add<ShapeOfCastExtentTensor, ShapeOfWithTensor>(context);
 }
 
+LogicalResult mlir::shape::ShapeOfOp::inferReturnTypes(
+    MLIRContext *context, Optional<Location> location, ValueRange operands,
+    DictionaryAttr attributes, RegionRange regions,
+    SmallVectorImpl<Type> &inferredReturnTypes) {
+  if (operands.size() != 1)
+    return failure();
+  if (operands[0].getType().isa<ValueShapeType>())
+    inferredReturnTypes.assign({ShapeType::get(context)});
+  else {
+    auto shapedTy = operands[0].getType().cast<ShapedType>();
+    int64_t rank =
+        shapedTy.hasRank() ? shapedTy.getRank() : ShapedType::kDynamicSize;
+    Type indexTy = IndexType::get(context);
+    Type extentTensorTy = RankedTensorType::get({rank}, indexTy);
+    inferredReturnTypes.assign({extentTensorTy});
+  }
+  return success();
+}
+
+bool mlir::shape::ShapeOfOp::isCompatibleReturnTypes(TypeRange l, TypeRange r) {
+  if (l.size() != 1 || r.size() != 1)
+    return false;
+  if (l == r)
+    return true;
+
+  Type lhs = l.front();
+  Type rhs = r.front();
+
+  if (!lhs.isa<ShapeType, ShapedType>() || !rhs.isa<ShapeType, ShapedType>())
+    return false;
+
+  if (lhs.isa<ShapeType>() || rhs.isa<ShapeType>())
+    // Shape type is compatible with all other valid return types.
+    return true;
+
+  if (succeeded(verifyCompatibleShapes({lhs, rhs})))
+    return true;
+  return false;
+}
+
 //===----------------------------------------------------------------------===//
 // SizeToIndexOp
 //===----------------------------------------------------------------------===//
 
 OpFoldResult SizeToIndexOp::fold(ArrayRef<Attribute> operands) {
   // Constant values of both types, `shape.size` and `index`, are represented as
   // `IntegerAttr`s which makes constant folding simple.
   if (Attribute arg = operands[0])

mlir/test/Dialect/Shape/invalid.mlir

func @shape_of(%value_arg : !shape.value_shape,
                %shaped_arg : tensor<?x3x4xf32>) {
   // expected-error@+1 {{if at least one of the operands can hold error values then the result must be of type `shape` to propagate them}}
   %0 = shape.shape_of %value_arg : !shape.value_shape -> tensor<?xindex>
   return
 }
 
 // -----
 
+func @shape_of_incompatible_return_types(%value_arg : tensor<1x2xindex>) {
+  // expected-error@+1 {{'shape.shape_of' op inferred type(s) 'tensor<2xindex>' are incompatible with return type(s) of operation 'tensor<3xf32>'}}
+  %0 = shape.shape_of %value_arg : tensor<1x2xindex> -> tensor<3xf32>
+  return
+}
+
+// -----
+
 func @rank(%arg : !shape.shape) {
   // expected-error@+1 {{if at least one of the operands can hold error values then the result must be of type `size` to propagate them}}
   %0 = shape.rank %arg : !shape.shape -> index
   return
 }
 
 // -----