do not make inferences with the same source\target pair multiple times

[vladima]

partial fix for #7081

var a = new Rx.Subject<string>();
var b = new Rx.Subject<Immutable.Set<number>>();
var c1 = a.withLatestFrom(b, x => x); // 1

Total compilation time for an original repro case (above) has dropped from +20 minutes to 1.5s (check time: ~0.9s ). However caching is local to every inference session so it still scales pretty badly

Number of times line (1) appear in the code	Check time (s)
1	0.93
5	1.12
10	1.37
20	1.79
40	2.75
80	4.54
160	8.28
320	15.59

Note: I'm not sure how to write a self-contained test for this issue without including rx.all.d.ts and immutable.d.ts

[vladima]

as a subsequent step we can set the maximum depth of nesting when we explore the structure of types during type inference. This approach seems to be promising and it scales better - these are numbers for the same test with max depth = 5

Number of times line (1) appear in the code	Check time (s)
1	0.81
5	0.88
10	0.93
20	1.08
40	1.27
80	1.72
160	2.69
320	4.49

However before moving forward I'd like to verify that setting max depth does not lead to serious regressions in type inference in our real world code test suite.

[DanielRosenwasser]

LGTM, others should take a look

[mhegazy]

👍

[sandersn]

Does this address the checking of typeof C vs instantiated typeof C in a generic static function? If so, you can use the test from #7138, at 7a48fb2.

[vladima]

no, #7097 is not fixed with this change

[sandersn]

Oh well. :( 👍, but with the caveat that my grasp on the caching mechanism is not too strong.