The ::mlir::Op class does not have a size of 8 when targeting the MSVC ABI caused by a longstanding bug of MSVC.
This may slow down the speed of MLIR Pass on Windows. See the example code with O2 optimization:
and the example code with O0 optimization:
The example shows that the size of mlir::Op<Traits...> without optimization is 72 (no_opt_size).
The optimized one with __declspec(empty_bases) is 8 (opt_size). The ::mlir::Op also affects the function call without inlining.
But I did not see much workaround in the LLVM project, other than this commit.
libc: Fix std::tuples EBO when targeting the MSVC ABI.
Should we add the attribute __declspec(empty_bases) for the ::mlir::Op class?
Myriad Dreamin