About the operation and its builder

MLIR
1

This is my td file contents:

def xxxOp : xxxOp<"xxx", [RecursiveMemoryEffects,
      SingleBlockImplicitTerminator<"YieldOp">,
      ParentOneOf<["func::FuncOp", "AffineForOp"]>]> {
   ...
  let results = (outs Variadic<AnyType>:$results);
  let regions = (region SizedRegion<1>:$body);
  let assemblyFormat = "attr-dict-with-keyword ( `:` type($results)^ )? $body";

  let hasVerifier = 1;
  let hasCanonicalizer = 1;


  let extraClassDeclaration = [{
    /// Get the terminator yield op.
    YieldOp getYieldOp();
  }];
}

My LLVM is 17.0.0. I have some problems with builder or rewiter. Here’s how I use them.

auto new =
          rewriter.create<xxOp>(op.getLoc(), ValueRange());

auto new = builder.create<xxOp>(loc,  returnValues);

I get an error when compiling：

error: no matching function for call to ‘mlir::xxx::Op::build(mlir::OpBuilder&, mlir::OperationState&, mlir::ValueRange)’
  489 |     OpTy::build(*this, state, std::forward<Args>(args)...);

If I modify it to the following：

auto new = builder.create<xxOp>(loc,  returnType, returnValues);

I don’t have problems when compiling, but I get these problems with mlir-opt :

:build(mlir::OpBuilder&, mlir::OperationState&, mlir::TypeRange, mlir::ValueRange, llvm::ArrayRef<mlir::NamedAttribute>): Assertion `operands.size() == 0u && "mismatched number of parameters"' failed.

I’m not sure what’s wrong and I’m looking for suggestions.

2

OOC have you switched your dialect to use properties?

3

Have you checked the available codegen’ed build()'s?
e.g., Tosa ops : $build/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.cpp.inc

void AbsOp::build(::mlir::OpBuilder &odsBuilder, ::mlir::OperationState &odsState, ::mlir::Type output, ::mlir::Value input1) {
...
}
void AbsOp::build(::mlir::OpBuilder &odsBuilder, ::mlir::OperationState &odsState, ::mlir::TypeRange resultTypes, ::mlir::Value input1) {
...
}
void AbsOp::build(::mlir::OpBuilder &, ::mlir::OperationState &odsState, ...

I’m not sure exactly how they’re generated but you can also add your own version of build() wrapping the existing version of build() in yourDialectOps.cpp as per your need.

4

I am a little confused, are you trying to say attributes, I have added to my dialect.

5

I actually wrote the custom builder in the td file, but I may need to manually rewrite it in the DialectOps.cpp . It’s worth mentioning that I’ve seen a similar format in someone else’s project file, but the pj compiles and executes fine, which is the point of my confusion.
Thanks.

6

I don’t think it’s necessary to implement in .cpp if you have a custom build in .td
But the error message suggests there’s something wrong in calling it (probably argument number/type mismatch)
You may want to check the generated code in .cpp.inc (I’ve pointed wrong path in my previous comment, it’s under $build/tools/mlir/…)

For example,

mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td

def SparseTensor_ForeachOp : SparseTensor_Op<"foreach",
    [SingleBlockImplicitTerminator<"YieldOp">]>,
     Arguments<(ins AnyTensor:$tensor,
   ...

  let builders = [
    OpBuilder<(ins "Value":$tensor, "ValueRange":$iterArgs, "AffineMapAttr":$order,
      "function_ref<void(OpBuilder &, Location, ValueRange, Value, ValueRange)>")>,
    OpBuilder<(ins "Value":$tensor, "AffineMapAttr":$order,
      "function_ref<void(OpBuilder &, Location, ValueRange, Value, ValueRange)>":$bodyBuilder),
    [{
      build($_builder, $_state, tensor, ValueRange(), order, bodyBuilder);
    }]>,
    OpBuilder<(ins "Value":$tensor,
      "function_ref<void(OpBuilder &, Location, ValueRange, Value, ValueRange)>":$bodyBuilder),
    [{
      build($_builder, $_state, tensor, ValueRange(), nullptr, bodyBuilder);
    }]>,
    OpBuilder<(ins "Value":$tensor, "ValueRange":$iterArgs,
      "function_ref<void(OpBuilder &, Location, ValueRange, Value, ValueRange)>":$bodyBuilder),
    [{
      build($_builder, $_state, tensor, iterArgs, nullptr, bodyBuilder);
    }]>
  ];

generates
build/tools/mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.cpp.inc

void ForeachOp::build(::mlir::OpBuilder &odsBuilder, ::mlir::OperationState &odsState, Value tensor, AffineMapAttr order, function_ref<void(OpBuilder &, Location, ValueRange, Value, ValueRange)> bodyBuilder) {
      build(odsBuilder, odsState, tensor, ValueRange(), order, bodyBuilder);
}
void ForeachOp::build(::mlir::OpBuilder &odsBuilder, ::mlir::OperationState &odsState, Value tensor, function_ref<void(OpBuilder &, Location, ValueRange, Value, ValueRange)> bodyBuilder) {
      build(odsBuilder, odsState, tensor, ValueRange(), nullptr, bodyBuilder);
}
void ForeachOp::build(::mlir::OpBuilder &odsBuilder, ::mlir::OperationState &odsState, Value tensor, ValueRange iterArgs, function_ref<void(OpBuilder &, Location, ValueRange, Value, ValueRange)> bodyBuilder) {
      build(odsBuilder, odsState, tensor, iterArgs, nullptr, bodyBuilder);
}
void ForeachOp::build(::mlir::OpBuilder &odsBuilder, ::mlir::OperationState &odsState, ::mlir::TypeRange results, ::mlir::Value tensor, ::mlir::ValueRange initArgs, /*optional*/::mlir::AffineMapAttr order) {
  ... }

void ForeachOp::build(::mlir::OpBuilder &, ::mlir::OperationState &odsState, ::mlir::TypeRange resultTypes, ::mlir::ValueRange operands, ::llvm::ArrayRef<::mlir::NamedAttribute> attributes) {
  ... }
7

Yes, thanks for your reply.

8

OK, fine. The problem is with the LLVM/MLIR version, i was using previously gcc, but switching to clang resloved it.