Re: [RFA] New testcase to evaluate Fortran substring expression

[Wu Zhou <woodzltc at cn dot ibm dot com>]

On Thu, 14 Jul 2005, Daniel Jacobowitz wrote:

> First of all, why do you need four operators?  I'd just use one, and
> push some sentinel value for the empty boundary case.  I'm not sure if
> you can use NULL in this context, but if you can, that'd work nicely.

I ever tried to use NULL, but it can't distinguish between ":exp" and 
"exp:".  But I am thinking of adding one longcst into operator 
OP_F90_RANGE to indicate the range type, 0 refer to ":",  1 refer to 
"exp:", "-1" means ":exp" and 2 means "exp:exp".  Or add two boolean 
to indicate whether that bound is empty or not.  What is your point on 
this.  I will try to do some tests soon.

> > ! arglist :      subrange
> > !                         { }
> >   	;
> >      
> 
> IIRC empty actions can just be skipped.

Do you means that I need to delete this empty action?  I see others keep 
this.  So I did this too.
  
> > ! arglist	:	arglist ',' arglist %prec ABOVE_COMMA
> >   			{ arglist_len++; }
> >   	;
> >   
> 
> The old grammar was:
>    arglist: /* empty */
>             | exp
>             | substring
>             | arglist ',' exp
>    substring: exp ':' exp
> 
> So "1, 1" "1:1", "1:1,1", "1:1,1,1" were all valid arglists.  I don't
> know Fortran, but I'm guessing this is intended, for multidimensional
> arrays?

They are not valid expression for multi-dimensional array.  I guess it is 
might be the fault of the original author.  This grammar was adapted from 
the C parser. (You have a look at the header of f-exp.y for proof). In C 
parser, it is just used to expression Function argments list.  But in 
fortran, things are somewhat complicated: parentheses are also express 
array subscript, substring and also array section (the latter is the F90 
extension).  I guess the case is that the original author just add 
substring as a valid arglist and didn't notice that it might interpret 
"1:1,1", "1:1,1,1" and such as valid too.

> 
> The new grammar is:
>   arglist : /* empty */
>             | exp
>             | subrange
>             | arglist ',' arglist

My intention is to use this to allow such things like "1:3, 2", "1:3, 2, 
2:5" and so on.  It is somewhat awesome to have to use "1:3, 2:2" or 
"1:3, 2:2, 2:5". 

> 
>   subrange : exp ':' exp
>              | exp ':'
>              | ':' exp
>              | ':'
> 
> One big problem here: I would have assumed bison would have bumped up
> the number of conflicts reported here.  Maybe it doesn't because of the
> %prec.  But you probaby don't want arglist ',' arglist.  It has all
> sorts of quirks; for instance, it allows "," as an arglist with two
> elements.  If what you are trying to do is allow "1:1, 1:1" then you
> want:

Yes, I ever had some thought about that.  If you really care, I can 
drop my grammar, and accept yours.

> 
>   arglist : /* empty */
>             | exp
>             | subrange
>             | arglist ',' exp
>             | arglist ',' subrange
> 

Thanks very much for your kind comments.

Regards
- Wu Zhou