try_fold support

[epage]

Prerequisites

Here are a few things you should provide to help me understand the issue:

• Rust version : 1.55.0
• nom version : 7.0.0
• nom compilation features used: default

Split out of #1409

fold_many_* doesn't support coming across errors. When folding the elements of a TOML table, we need to error if keys are duplicated. I think it'd be useful to take inspiration from try_fold in the stdlib. This uses an unstable trait but we can for now constrain what types we can Try on until Try becomes stable.

[epage]

We should wait until #1393 so we can leverage it for only having 1-2 versions, rather than 4+

[Geal]

good idea, and yes, it'd make more sense after we get ranges

[Stargateur]

This uses an unstable trait but we can for now constrain what types we can Try on until Try becomes stable.

So I try...

And it very very very very very very very very very complicated.

1. I add a nightly feature
2. I go for fn try_fold => IResult<Input, Try<Output = Acc, Error> problem this make an result inside a result so I didn't push it
3. I go for fn try_fold => IResult<Input, Acc, Error> where F: FnMut(Acc, Output) -> Try, Try: std::ops::Try<Output = Acc, Residual = Error>, merging the error... problem big problem I let you read the new very short RFC

The problem is that all nom parser return an Result and the Residual type for impl Try for Result is Result<Infallible, E>. This make any generic instantly fail. The user can't use it cause

the trait `error::ParseError<_>` is not implemented for `Result<Infallible, error::Error<&str>>`

While maybe we could add it, I think this will further add problem. I think using Try would make sense if we return the Try inside try_fold. This mean maybe nom would completely rewrite its error type, for example a NomResult that implement Try allowing a lot of thing.

So, for now I think simple require the function to return a Result<Acc, Err> is much better. (Err of nom allowing to return failure error for allocation failure for example)

[epage]

So, for now I think simple require the function to return a Result<Acc, Err> is much better. (Err of nom allowing to return failure error for allocation failure for example)

This was my intention, at least until Try became stable, and then we could re-evaluate. Sounds like that evaluation has happened now :)