On Mon May 13 19:14:17 2024 +0000, Gabriel Ivăncescu wrote:
They could potentially be just indices, the association between name
and ID may be easily done on MSHTML side (it's already doing that now). I don't think I follow. The name is used for jscript prop lookup, not mshtml. We absolutely need it for hash and standard name lookup. It's not an internal thing. PROP_PROXY doesn't care what name it has, it only stores the DISPID (index basically) to access it from mshtml. But looking it up **does** require name, because the prop has a name, so we do need dispex_prop_t, just like builtins. To be fair I'm talking about **being accessed by name via jscript**, nothing to do with mshtml internals. Think of window props, for example. I'm not talking about simple dynamic props on the window, which can be stored in jscript as simple values (and will be, later though, not in my branch yet, as I said I don't want to overcomplicate it). Instead, think of the other "global_props" (from mshtml's meaning of it) on the window, such as other scripts, elements, etc. Those need PROP_PROXY to get accessed…
I think I can live with it, but note that the requirement to create a
function object instance for all MSHTML builtin functions that are used is not exactly optimal. It would be possible to optimize that in the future, hopefully without full rewrite of the interface. Just like we do for jscript builtin functions. We can save that for later, but I believe it's as optimal as builtins right now. That is, they get allocated on demand (when needed), but otherwise we need them as functions, because even builtins had to be converted from builtins into actual functions at some point. Only the "value prop" builtins still remain as such (it's same with PROP_PROXY).
Sure, although... it's by no means something for now, but I could
imagine getting rid of `PROP_BUILTIN` and use the same 'external' properties for them instead, just like other external props (MSHTML and indexed props in the current proposal). How would that work? Builtins have a name, so we can't just use the same as indexed props because such name needs to have a DISPID mapping to it and be looked up by hash. We could use an extra layer of mapping for builtins of course, but I think that's overkill and likely slower (instead of just one hash lookup, now it's two lookups). It also complicates it because, for instance, builtins can be deleted, and then we have to actually implement every prop behavior on them instead of relying on dispex_prop_t facilities. Right now if they get deleted, the prop remains as PROP_DELETED and so can be easily replaced by something else. But yeah, maybe it's worth it way later though, we'll need measurements I guess.
After rebasing more of my patches, I found a problem with the IID approach, when it comes to weird prototype chains. For example the HTMLCurrentStyle shares an "abstract" common prototype with HTMLStyle, even though some of those methods don't work on it. But some are exposed in multiple interfaces, for instance the `zoom` prop setter (not getter).
The IID would have the IHTMLStyle3 interface (which is correct), but when looking the "zoom" prop up on HTMLCurrentStyle via DISPID, it would get the one from CSSStyleDeclaration, which has different behavior. Yes, it's weird as hell, and my explanation isn't that great either.
Anyway, the IID isn't that useful, mshtml already queries the IID for the respective function.
Instead I'm now using a mixed approach where I have a new method (but a normal method in the vtbl like PropInvoke rather than a function pointer callback) that uses a "func_ctx" which is basically the `func_info_t` (it's opaque to jscript, it doesn't care). This way there's zero ambiguity as to which function is actually used, and the prototype chain fiasco is cleanly solved. All tests pass now.
Anyway, this is not very related to this MR, is there something I can do to get this moving?
-
Gabriel Ivăncescu (＠insn)