Circular precedence

[Valentin Tolmer]

In following the patch suggested about partial order precedence, I made 
a test case with a circular precedence, i. e. c > b > a > c
As far as I know, such a precedence relationship was not only impossible 
to express directly with only precedence (%left etc...), but it was 
completely impossible, even using something like
e:
    e * e
|   f

f:
    f + f
|   NUM

(this represents * > +)

With the new system, it is entirely possible and gives an interesting 
result. As there is currently no check for circularity and (yet) no 
transitivity (although non-transitional relationships will always be 
possible), a naive resolution for this grammar
%token NUM

%gprec {
  %left a
  %left b
}

%gprec { %left c }

%precr a > c
%precr c > b

%%

e:
  e a e
| e b e
| e c e
| NUM

expresses these parses
1 a 2 a 3 ->  (1 a 2) a 3         //normal %left associativity
1 a 2 b 3 ->  1 a (2 b 3)         //b > a
1 a 2 c 3 ->  (1 a 2) c 3         // a > c
1 c 2 b 3 ->  (1 c 2) b 3         // c > b
1 b 2 a 3 ->  (1 b 2) a 3         // b > a

The question is, should this be possible? Is it needed? Is it a 
desirable behavior?
For my part, I see no harm in keeping it, but maybe preventing users 
from creating more complicated cycles inadvertently would be better... 
(Maybe a warning with an %expect-cycle 1 like we have %expect-rr ?)

Tell me what you think!