Updates to make ffi-napi compatible with an updated ref-napi

[ggryschuk]

Hi. As noted under the ref-napi issues list I've been wanting to update to use the ref-napi package but it's performance (and memory usage) was significantly slower (and larger) than the ref(NaN) version as others have observed. The ref-napi performance issues also impact ffi-napi.

The good news is I've identified the updates that need to be made to ffi-napi to make it properly work with a ref-napi with the 'performance updates' applied (including an additional update to ref-napi I identified as necessary, see my comments here 72 and here 76).

The updates necessary to ffi-napi are to code that expects that a pointer is a '0-length buffer'. I haven't found anything in the documentation that actually indicates that this is or should be true, and from what I can tell a pointer should be a '1-length buffer' (to hold the pointer itself). With the changes to ref-napi for performance and memory usage a pointer explicitly becomes a '1-length Buffer' which I believe is it's correct behavior (I do have to admit to being confused as to how the previous behavior actually worked in so much as storing a pointer in a Buffer that became 'zero length' on return from the C/C++ code).

Long story short, ffi-napi calls 'WrapPointer' of the ref-napi package (at the C/C++ code level) without specifying a size, since ref-napi now makes a pointer Buffer a 1-length Buffer the existing ffi-napi package won't work. The fix is to add a size of '1' to these WrapPointer calls, here's an example:

original line:

o["dlopen"] = WrapPointer(env, dlopen);

Updated line

o["dlopen"] = WrapPointer(env, dlopen,1);

I'm attaching a diff of the changes to be made to src/callback_info.cc, src/ffi.cc, test/ffi_test.cc, test/dynamic_library.js (to account for a test that assumed a pointer was a 0 length buffer).

With the updates to ref-napi & these updates the ffi-napi package passes all it's mocha tests.

Equally importantly (to me anyway) is that I've tested these changes to my own project & they work great. This was a limited but full test of my code that makes extensive use of ffi-napi, ref-napi & ref-struct-di, the performance was easily on par with the equivalent NaN packages, potentially slightly better.

I hope that someone will apply these patches (or equivalent ones in case I've entirely misunderstood how the ref-napi package should create a pointer Buffer) to the official ref-napi & ffi-napi packages so that others can test them.

Thanks.

PS. My apologies if this isn't correct usage of this forum. If there was a better way to submit this please advise.

node-ffi-napi-diff.txt