flutter_test_performance regressed by ~3x

[liyuqian]

The regression clearly started from 9d7116c

A bisect for engine commits is needed.

[liyuqian]

CC test author @Hixie to see if there's any insight.

[jason-simmons]

Bisected this to dart-lang/sdk@53ac4d5

@jensjoha

[jensjoha]

It became slow with dart-lang/sdk@53ac4d5 because the nnbd mode is now actually set for the platform (it must have not been before). It is compiled as nnbd weak which is not allowed to be mixed with nnbd disabled (which is what the other stuff is compiled as).

I'm not totally sure if it actually should be allowed --- @johnniwinther or @mraleph might know...

(Also it seems that we haven't thought about this situation where we load a single dill that's not agnostic and compile against that in another mode --- probably we should; again, @johnniwinther might have thoughts?)

Update: I accidentally spelled @mraleph wrong so I pinged a wrong person. Sorry about that.

Update #2: Let me try the explanation of why it becomes slow again. Before it's fast because we initialize from dill. After it's slow because as we try to initialize from dill we say the dill is invalid because it's in a "mixed mode nnbd state", and we thus doesn't initialize from dill but actually compiles from scratch.

[liyuqian]

@jensjoha : shall we revert the commit, and reland later once we figured out either a fix, or that we should rebaseline the benchmark to allow this regression?

[Hixie]

We should definitely not regress test performance 3x, it's already 10x slower than it was a few years ago.

[jensjoha]

We will allow mixing disabled and weak. That should fix the issue.

[jensjoha]

This should now be fixed in dart, meaning it should go back to normal when dart is rolled.