Do we want to specifically ignore these by list all 6 files here?

how about including either the original file or the modified file from cmake?
could we copy the file to the build directory instead of within the source tree? that may support multiple builds better.

I thought about and tried that also, that approach will be modifying the onnxruntime_providers.cmake a lot to remove and add new files to many source lists for cpu and cuda eps, also need some change in onnxruntime_unittests.cmake for training ops. This is hard to maintain and also require same amount of work if we want to add reduced kernel for new EPs.
The current impl is not ideal, but it is relatively simple to implement and maintain.
We can certainly improve this later

ok. I assume that was a response to my first question.

how about the second:
could we copy the file to the build directory instead of within the source tree?

Actually create ***_reduced.cc/inc in build folder also requires a lot of changes to the onnxruntime_providers.cmake
I have a working version here
https://github.com/microsoft/onnxruntime/tree/gwang-msft/better_reduced_build_test

I still prefer the current simple impl despite its limitation

We also need to be careful about it working with different styles of build. IIRC it's possible to run the reduction script standalone prior to running build.py or cmake directly, so we need to allow for that. If I ran the script, it updated something in the build dir, and then I did a 'clean' it's somewhat opaque that I just lost the reduction info and need to run the script again.

Would it be relatively simple for the cmake to remove either the original or the reduced file from the globbed list (e.g. onnxruntime_provider_srcs) based on whether it's a reduced build or not? that way we don't need the ifdefs. something like:

if (onnxruntime_REDUCED_OPS_BUILD)
  list(REMOVE_ITEM onnxruntime_providers_srcs ${ONNXRUNTIME_ROOT}/core/providers/cpu/cpu_execution_provider.cc)
else()
  list(REMOVE_ITEM onnxruntime_providers_srcs ${ONNXRUNTIME_ROOT}/core/providers/cpu/cpu_execution_provider_reduced_ops.cc)
endif()

one idea:
for the files involved in reduced op types, we could define a root path variable in cmake, say ${ONNXRUNTIME_OP_REDUCTION_ROOT}. when no op reduction is done, that is simply ${ONNXRUNTIME_ROOT} (maybe we'd need a similar one for training). otherwise, it is a path to another location (I'd prefer one in the build directory) that the modified files are in. regardless of root, we could keep the root-relative parts of the path the same.
then we can refer to ${ONNXRUNTIME_OP_REDUCTION_ROOT} in the cmake files, add it as an include directory, and keep any relative references in the source files unchanged (e.g., in op_kernel_type_control.h).

We also need to be careful about it working with different styles of build. IIRC it's possible to run the reduction script standalone prior to running build.py or cmake directly, so we need to allow for that. If I ran the script, it updated something in the build dir, and then I did a 'clean' it's somewhat opaque that I just lost the reduction info and need to run the script again.

perhaps adding a (required?) --build_dir option to the reduction script would make it clearer.

If you look at my other branch
https://github.com/microsoft/onnxruntime/tree/gwang-msft/better_reduced_build_test

I added an optional output_path, we can change this to mandatory

I still prefer my current impl which is very simple, which is an improvement over what we have now and will work for most cases.
The only issues are that it generates the reduced_op code files in src folder and use ifdef
Generating reduced_op code files in build folder or other folder is a better idea but will need more consideration for many cases.

We can log a future work item to further improve this