Re: consistent states and error transitions

[Joel E. Denny]

On Sun, 27 May 2007, Paul Eggert wrote:

> "Joel E. Denny" <address@hidden> writes:
> 
> > I'd like to commit the following.  I'd appreciate a second pair of eyes to 
> > be sure my assumptions are correct.
> 
> The change looks OK to me, not that I've thought it through as
> carefully as you have.

Thanks for reviewing it.

> However, shouldn't there also be a change to
> the comment for 'consistent'?  Currently it says:
> 
>   /* Nonzero if no lookahead is needed to decide what to do in state S.  */
>   char consistent;
> 
> Doesn't one in general need lookahead to decide what to do in an error
> state?

For states that shift the error token, I believe it makes sense the way 
it's worded.  If such a state has no reduce actions, then lookahead sets 
are not possible, and the state is considered consistent.  If such a state 
has reduce actions, then the state needs lookahead sets and is considered 
inconsistent.  Does that sound reasonable?

The only flaw I see with the wording is that it claims that inconsistent 
states always need lookaheads.  Consider a state that doesn't have a shift 
on the error token, that does have at least one other shift, and that has 
only one reduce action.  That reduce action becomes the default rule, so 
it doesn't really need lookaheads, but the state is still considered 
inconsistent because it has both reduce and shift actions.

How does the following sound?

  /* If non-zero, then no lookahead sets on reduce actions are needed to 
     decide what to do in state S.  */
  char consistent;

Or did I miss the point?