More fine grain control of precedences

[theKidOfArcrania]

Hi I'm having trouble trying to disambiguate this fairly simple grammar (note that this is only a small chunk of a larger language):

grammar;

pub ETop = <E>;

E: String = {
  #[precedence(level="0")]
  var => <>,
  "let" <v: var> "=" <e1: ETop> "in" <e2: ETop> =>
    format!("(let var = {e1} in {e2})"),

  #[precedence(level="1")] #[assoc(side="left")]
  <e1:E> ";" <e2: E> => format!("({e1} ; {e2})"),
}

var: String = {
  r"[A-Za-z]" => <>.into(),
}

This would result as the following shift-reduce conflict:

    The following symbols can be reduced in two ways:
      "let" var "=" ETop "in" E ";" E0

    They could be reduced like so:
      "let" var "=" ETop "in" E      ╷ ";" E0
      │                       └─ETop─┤      │
      ├─E0───────────────────────────┤      │
      ├─E────────────────────────────┘      │
      └─E───────────────────────────────────┘

    Alternatively, they could be reduced like so:
      "let" var "=" ETop "in" E ";" E0
      │                       ├─E────┤
      │                       └─ETop─┤
      └─E0───────────────────────────┘

However we also can't place the let ... in ... production rule as a lower precedence to everything else because then we wouldn't be able to parse inputs such as b; let a = b in c. In essence, we want to have let be part of a set of high precedence keywords, but also let the trailing expression have low precedence when reducing.

The solution (using conditional macros/explicitly handling no-short-if case) in #386 does not work as both solutions (the linked example code) nor the linked java grammar has the issue where we have a keyword "let" with a high priority yet have the trailing expression as low priority (i.e. a mixed context-dependent precedence) (note ternary in java tactfully does not have an "if" to denote the beginning of a ternary expression, so it can just be placed as a low priority operator).

A solution in yacc has been to mark the let keyword as high precedence, but mark the whole production as a low precedence, ex:

%token LET IN
%token<str> VAR

%nonassoc P_LET
%left ';'
%nonassoc LET

%%

expr: VAR { }
    | LET VAR '=' expr IN expr %prec P_LET { }
    | expr ';' expr { }
    ;

In essence, we are saying to prefer shift over reduce if we are in the let ... in ... production. This seems to be similar to issue #620 but more generalized.

Note this was also briefly mentioned in #67, though was never actually addressed.