> First, the proposed merge requests adds a ton of conditional code paths that are impossible to properly review now or even maintain over time. Since there is no proper distinction in the source code between what is supposed to be shareable+readonly and exclusive+mutable, There is no way by looking at the code to determine the correctness of the change, this will make improving library internals even more difficult in the future. It is very likely that this feature is gonna break in subtle ways that are very hard to detect beforehand. Given the scope of the library's API, I don't think it is possible to come up with an extensive regression test that covers every possibility. I am convinced this will not be maintainable over time. Sorry.
I get your point and have a few options in mind to overcome this. But it would require some structural changes. That's why I was deliberately being quiet about that. So basically, we need to organize the fields in such a way that they reflect their traits. It can be having a prefix describing the trait or organizing like the following:
typedef struct ABC {
struct {
/* Mutable Properties. */
} mutable;
struct {
/* Lazy Properties. */
} lazy;
/* Immutable Properties. */
FT_UInt first;
FT_UInt second;
} ABC;
Let me know if this sounds interesting, then we can further refine it.
> The locking requirements do not only apply to the input stream but to any data that is loaded or created on demand. A lot of these have been removed to make the library "more thread-safe" (it really isn't). And since the FreeType API was designed to use mutable state very intentionally (this saves a lot of memory compared to the use of immutable data structures, which was important for embedded systems). There is no clear way for a client to determine whether a given API will mutate the state, so the only safe option for a client is to lock on every API call.
Again, I agree with this. But I have shared the problem statement and how immutability can overcome that. Definitely the safe option would be to acquire the lock on every API call but we can document the APIs that would need it. I was only referring to associate the lock with the input stream, not necessarily acquiring it in read operation.
> Finally, calling these instances "clones" or "copies" is misleading. A best they should be called "dependent faces" or "child faces", but the fact that they complicate the lifecycle management of many objects is a problem in itself.
We can also consider to call it a "derived" face and name the function "FT_Derive_Face" instead. We can also keep the list of child faces in the root face to manage their life cycle. A child face would retain the parent face and remove itself from the list on deallocation.
> An FT_Clone_Face() API that takes an input FT_Face instance and returns a new instance that has its own lifecycle (except that they will be children of the same FT_Library, and use the same FT_Memory allocator).
I have shared an alternate above.
> For the input stream, a way to either clone it, or share it safely between instances is needed, and should be provided as an argument to the function in some way. We could change the stream implementation used internally by the library to make this easier, or we could require the client to use FT_Open_Face() with a user-provided shareable stream for FT_Clone_Face() to work.
It seems a little complicated to me. Besides, some OS don't allow many instances of the same file.
> The initial implementation would simply re-open the face with the new stream, inefficient but completely safe. But this opens the door to identifying read-only data in the source face that can be copied directly to the clone, saving all the work required to load/validate/compute it.
I have already identified the traits of variables. They are written alongside each copy operation in clone implementations.
> Note that I wrote "copy" above, because for efficient and safe read-only sharing, atomic reference counting is required, which is not part of C99, hence not portable. However, it can be introduced as a separate step by defining the right abstractions (types and macros to operate on them). Essentially, we need the equivalent of std::shared_ptr<>, or the intrusive versions of it where the ref-count is at the start of the shared object, but in C99 instead. For platforms that do not support threads, just do non-atomic refcounts.
I also had this in mind but I was trying my level best to keep things simple and additive only. I was thinking more of requiring atomic integer implementation from the client side because it's cheaper than a lock.
If we decide to go this way then this would need to be catered first, I guess.