[C++][Compute] Remove redundant cast kernels

[andishgar]

Describe the enhancement requested

In Arrow's cpp/src/arrow/compute/kernels/scalar_cast_string.cc, there are several cast kernels registered for cases where the input and output types are the same, for example:

arrow/cpp/src/arrow/compute/kernels/scalar_cast_string.cc

Lines 643 to 658 in 73742e8

	// Fixed -> Fixed
	template <typename O, typename I>
	enable_if_t<std::is_same<I, FixedSizeBinaryType>::value &&
	std::is_same<O, FixedSizeBinaryType>::value,
	Status>
	BinaryToBinaryCastExec(KernelContext* ctx, const ExecSpan& batch, ExecResult* out) {
	const CastOptions& options = checked_cast<const CastState&>(*ctx->state()).options;
	const int32_t in_width = batch[0].type()->byte_width();
	const int32_t out_width =
	checked_cast<const FixedSizeBinaryType&>(*options.to_type).byte_width();
	if (in_width != out_width) {
	return Status::Invalid("Failed casting from ", batch[0].type()->ToString(), " to ",
	options.to_type.ToString(), ": widths must match");
	}
	return ZeroCopyCastExec(ctx, batch, out);
	}

arrow/cpp/src/arrow/compute/kernels/scalar_cast_string.cc

Lines 472 to 487 in 73742e8

	template <typename O, typename I>
	enable_if_t<is_binary_view_like_type<I>::value && is_binary_view_like_type<O>::value,
	Status>
	BinaryToBinaryCastExec(KernelContext* ctx, const ExecSpan& batch, ExecResult* out) {
	const CastOptions& options = checked_cast<const CastState&>(*ctx->state()).options;
	const ArraySpan& input = batch[0].array;
	if constexpr (!I::is_utf8 && O::is_utf8) {
	if (!options.allow_invalid_utf8) {
	InitializeUTF8();
	ArraySpanVisitor<I> visitor;
	Utf8Validator validator;
	RETURN_NOT_OK(visitor.Visit(input, &validator));
	}
	}

arrow/cpp/src/arrow/compute/kernels/scalar_cast_string.cc

Lines 772 to 780 in 73742e8

	void AddBinaryToFixedSizeBinaryCast(CastFunction* func) {
	AddBinaryToFixedSizeBinaryCast<StringType>(func);
	AddBinaryToFixedSizeBinaryCast<StringViewType>(func);
	AddBinaryToFixedSizeBinaryCast<BinaryType>(func);
	AddBinaryToFixedSizeBinaryCast<BinaryViewType>(func);
	AddBinaryToFixedSizeBinaryCast<LargeStringType>(func);
	AddBinaryToFixedSizeBinaryCast<LargeBinaryType>(func);
	AddBinaryToFixedSizeBinaryCast<FixedSizeBinaryType>(func);
	}

However, casts between identical types appear to be handled earlier in cast.cc:

arrow/cpp/src/arrow/compute/cast.cc

Lines 101 to 127 in 73742e8

	if (args[0].type() && args[0].type()->Equals(*cast_options->to_type)) {
	// Nested types might differ in field names but still be considered equal,
	// so we can only return non-nested types as-is.
	if (!is_nested(args[0].type()->id())) {
	return args[0];
	} else if (args[0].is_array()) {
	ARROW_ASSIGN_OR_RAISE(std::shared_ptr<ArrayData> array,
	::arrow::internal::GetArrayView(
	args[0].array(), cast_options->to_type.owned_type));
	return Datum(array);
	} else if (args[0].is_chunked_array()) {
	ARROW_ASSIGN_OR_RAISE(
	std::shared_ptr<ChunkedArray> array,
	args[0].chunked_array()->View(cast_options->to_type.owned_type));
	return Datum(array);
	}
	}
	Result<std::shared_ptr<CastFunction>> result =
	GetCastFunction(*cast_options->to_type);
	if (!result.ok()) {
	Status s = result.status();
	return s.WithMessage(s.message(), " from ", *args[0].type());
	}
	return (*result)->Execute(args, options, ctx);
	}
	};

As a result, it seems these kernels may never be invoked.

Component(s)

C++

[andishgar]

@pitrou , could you verify my observation? If I'm correct, I should probably check whether similar redundant kernels exist for types other than strings as well.

[pitrou]

You could try disabling those casts and see if everything still works fine.

Some random observations:

1. FixedSizeBinary is parametric, for example FixedSizeBinary(2) and FixedSizeBinary(3) are different types
2. binary_view and utf8_view are different types