[release/8.0] Fix to #31961 - Projection filtering on second-level reference could not be translated

[maumar]

Problem was that optimizing visitor was looking for a very specific pattern (member access over conditional), but in this scenario conditional is wrapped around Convert node. Fix is to recognize this case also, strip convert around the conditional and apply it around non-null portion instead, before applying the member access over it.

Fixes #31961

Risk

Low - change is isolated to a narrow optimization (member access over conditional) that also uses type conversion (either from derived to base type or from non-nullable to nullable), and is a simple change (strip convert around convert node and apply it on the non-null part of it before applying the member access.

[maumar]

@roji could you take a look?

[ajcvickers]

@roji Can you take a look at this?

[roji]

LGTM, see comments below on the Convert node preservation.

Looking at this more generally, the optimization seems to make quite an effort to apply only when the conditional test is a comparison to null; but I'm not sure why that is - couldn't this be generalized?

i.e. instead of only applying to:

(a != null ? new { Member = b, ... } : null).Member

... couldn't this apply to any test expression:

(<arbitrary expression> ? new { Member = b, ... } : null).Member

... transforming it to:

<arbitrary expression> ? b : null

(this would obviously be outside the scope of this PR)

[maumar]

@roji we could try the general purpose optimization, although idk if we wouldn't be losing some context by doing it. This optimization is not mathematically correct on the Core level - we should be pushing down the member to both sides of the conditional expression. However, in case of x != null ? x.Member : null we recognize this as common patten to guard against null refs, so we take some liberty here. Either way, it's currently out of scope for sure.