Ion team, would you take a look at this, please? > From: Allen Wirfs-Brock <[email protected]> > Subject: Is dyntamic super binding really too expensive? > Date: March 28, 2013 12:32:33 PM CDT > To: Jason Orendorff <[email protected]> > Cc: Brendan Eich <[email protected]> > > ES6 currently specifies that any method that uses super has a static > reference to the "home" object that is the starting point (actually, minus 1) > of its super property lookups. This approach was taken to avoid adding the > overhead of an additional "where found" implicit argument to every function > call. However, it introduces various observable complications when such > functions are dynamically moved around by JS code. > > In the comments to https://gist.github.com/getify/5253319 I discussed this > and tried to defend the current design choice on the performance basis. > However, in the process it raised some doubts in my mind whether there really > is a significant perf issue with dynamic super binding. I can think of > various ways that at the implementation level I might be able to make that > cost approach 0. For example, if an actual property lookup is performed both > the actual this value and the "home" object of the resolved method should > already be available. In a register rich environment (x64, etc.) I can > imagine structuring the calling convention such that "home" is implicitly > passed in a register and only super-using callee's would need to capture it. > > Obviously, when aggressive interprocedural analysis and/or inlining is > performed the home inplicit argument overhead could be eliminated. The only > case I'm less sure about is when PIC based dispatching is performed. I'm not > up to speed on what the PIC look like for various modern engines but it seems > likely that a PIC needs to capture the "home" reference and that it also > could be passed as a 0 cost implicit parameter. > > So, here's my question. Can we convince ourselves that a typical JS engine > could reasonable approach 0 cost passing of a "where found" implicit > parameter to method calls. Even more concretely can you see how that might be > accomplished within Spider Monkey? If it's plausible, we might be able to > have a cleaner observable ES6 semantics for functions that reference super. > > What do you think? > > Allen >
_______________________________________________ dev-tech-js-engine-internals mailing list [email protected] https://lists.mozilla.org/listinfo/dev-tech-js-engine-internals
