This should be moved to a separate bench folder as we discussed in the binary ops PR review?

And separated into benchmarks and unit tests?

done

Should add more documentation/comments to the tests

done

@nsakharnykh Can you approve your review since this is resolved?

The interface is confusing to me. Why does it have column and to_replace parameters? What happens to column? What happens to to_replace? Why not just make a gdf_copy(out_column, in_column)?

Now I see. The meaning of "replace" is not clear. It's not a copy. The semantics are actually those of "find_and_replace_all": For each value in to_replace, find all instances of that value in column and replace it with the corresponding value in values. This should be made clear in the header documentation for the function. Consider changing the name to gdf_find_and_replace_all() or something like that...

We're adding google benchmark in this PR? Shouldn't that be orthogonal to gdf_replace, and therefore a separate PR?

Since this is the first benchmark I've seen added to libgdf...

Since no other PR depends of this, I put google benchmark here. But I could make another PR if it's relevant.

I don't really like calling this macro "WHEN" -- it's not descriptive. Something like REPLACE_CASE would be clearer. Then you would have

switch (column->type) {
    REPLACE_CASE(INT8);
    REPLACE_CASE(INT16);
    // etc.
}