$parse performance/style

[jbedard]

c074181 might be the only one worth merging to allow the Parser/Lexer objects to be GCed.

f4ea3c9 addresses a TODO although I didn't notice any major improvements (even though the removed wrapper method is often ~10% when profiling...). Unfortunately the wrapper method is still needed for the expressions returned from $parse, but use of getterFns within other expressions don't have the wrapper. This also required a minor error message change.

f4ea3c9 could be expanded in other ways though. Maybe allowing $parseed methods to return something such as a $$watchFn property that can be watched instead of the full expression being watched. The simple case would be returning the raw getterFn instead of the wrapper for simple foo.bar expressions. A more complicated case would be expressions such as myVar % 2 === 0 returning only the myVar getterFn to be watched (and then a special watch delegate to see if the full expression actually changed, which makes it more complicated...). Or only watching the input+args to filters etc. Not sure how common or useful those would be though...

The rest are mostly style/simplifying but I thought I'd leave them in there.

[btford]

@jbedard thanks for the PR! can you add a benchmark to https://github.com/angular/angular.js/tree/master/benchmarks to prove that these changes are beneficial?

[jbedard]

@btford I've added a new benchpress benchmark for execution of $parse()ed expressions (fd45e8a). It currently has tests for the different types of expressions (property, binary op, filter, ...). Basically it generates data (2000 rows right now), ng-repeats a bunch of expressions (12 per test right now, except object literals that are too slow) and then measures the digest (x50 right now) time.

I updated eb47e91 such that the wrapper function is now only needed for one-time binding so $parse()ed properties now benefit form this as well. (A wrapper is needed for one-time because the one-time :: is not actually part of parsing so both ::foo and foo now return the exact same getterFn result, but they have different $$watchDelegates so they need to be different functions returned from $parse).

Here are the numbers (ms) for the current tests in the benchmark (12 expressions x 2000 rows x 50 digests):

Test	Old	New	Explanation
simple properties	110	94	needs a few runs to stabilize but seems consistent
field accessors	193	178	$parsePathGetter was ~5% in a quick profile which matches the numbers
field indexors	1160	1160	$parsePathGetter was negligible compared to ensureSafeObject + $parseObjectIndex
prop/operators	206	186	$parsePathGetter was ~8% in a quick profile which matches the numbers
prop/filters	328	326	$parsePathGetter was negligible compared to $parseFilter + pipe operator + binaryFn
prop/literals/func	559	522	$parsePathGetter was mostly negligible compared to $parseFunctionCall + .apply + ensureSafeObject
prop/literals/{}	1474	1440	$parsePathGetter was negligible since it is mainly constants in the object, the .length seems to make the minor difference
prop/literals/[]	1003	979	(same as object literal)

[jbedard]

Thanks for looking into all of these and getting them in so fast! And note my comment regarding the changed tests/error-message if you haven't already (can avoid changing the tests/message if we want, see jbedard@cec3344)