[clang] [clang] Fix -Wdouble-promotion in C++ list-initialization (PR #159992)

[Marcel Jacobse via cfe-commits]

mjacobse wrote:

Yeah that makes sense to me. So I reverted this attempt of just checking if we are in some kind of ListInit, it does not work.

It looks like expressions like `double{f}` turn into a `CXXFunctionalCastExpr` just like `double(f)`, with the only difference that for `double{f}` the `CXXFunctionalCastExpr` gets an additional direct child `InitListExpr` with one initializer. So the implicit conversion analysis does skip over `CXXFunctionalCastExpr` as an explicit cast, but not over the `InitListExpr`, causing the warning. But it seems like at least for this purpose, `InitListExpr` basically belongs to the explicit cast still, so I added some code to skip it too in this case. And this seems to work as intended.

The implementation is not the prettiest though and it seems to rely heavily on this implicit assumption that `InitListExpr` as a direct child of `CXXFunctionalCastExpr` means that we have this special case of list-initialization. I'd be very happy about suggestions for how to improve this.

I also applied the test deduplication that you suggested and added some tests for classic function-style casts and the initialization examples that showed that the previous fix was broken. I can add further suggested tests involving classes and templates later if the new fix turns out to be the right idea.

I also introduced an alias for `long double` in the tests to write `LongDouble{f}` instead of `(long double){f}`. The latter did not do what I expected, instead it results in a C99 compound-literal construction (which does trigger `-Wc99-extensions`). These actually produce the `-Wdouble-promotion` warning too while they probably shouldn't, but I suppose that's something for a different issue.

https://github.com/llvm/llvm-project/pull/159992