[clang] [llvm] [mlir] [clang][APFloat] Print floating-point literals in the shortest round-trip form (PR #218471)

[Emery Conrad via cfe-commits]

conrade-ctc wrote:

Oh, wow, @AaronBallman, in adding a barrage of tests, I found some pre-existing I found 3 existing bugs in `APFloat`

> > Generally LGTM though I did not check the implementation of the algorithm very carefully. The results look correct to me in terms of the testing changes, but I think there are some tests we should add:
> > ```
> > double bar = 0x10.1p0;
> > ```
> > 
> > 
> >     
> >       
> >     
> > 
> >       
> >     
> > 
> >     
> >   
> > do we correctly round-trip that or do we still convert that to a decimal value instead of the hex literal?
> > Another question would be: do we have sufficient test coverage for some of the odder floating-point formats like double double or all the tiny floating point types for AI?
> 
> Good point. I'll add a barrage of corner case tests, compiling a list now, and I'll ping when they are pushed to the PR.

@AaronBallman, radix is not preserved in round-trip... that requires a bit more surgery (need radix field in AST nodes, etc). Okay with you to post an issue for that one, and follow-up with a separate PR, or do you think it's worth adding to this PR?

Also, glad you pressed on wider test coverage, other that the existing bug @efriedma-quic stressed in the comment above, I stumbled on another existing bug in `toString`... I'll add another issue for that one, and follow up with a separate PR to address, unless you think it makes sense to pack it all into this PR.

What do you think: add the issues and follow up with PRs, or get them into this one?

https://github.com/llvm/llvm-project/pull/218471