Re: range types, ptrdiff_t, signed integers for internals

[Paul Eggert]

On 12/12/19 12:50 AM, Bruno Haible wrote:

> Several programming languages have range types:

Pascal is where I first saw them.

> I would distinguish the range type [0 .. PTRDIFF_MAX] (which is an abstract
> concept) from the actual implementation (through the typedef). It could
> be represented by a signed type or an unsigned type; either is fine.

Part of the motivation for this change is that unsigned types are not 
"fine": they can't be checked for overflow via -fsanitize=undefined, and 
they are not optimized as well for index calculations and the like due 
to their wraparound overflow semantics.

> > Also, where would we put these new typedefs? dfa.h doesn't seem like the
> > right place.
>
> That's up to you. It could well be a separate .h file that is a module of
> its own (like verify.h and intprops.h are).

For now I was more conservative installed the attached, which limits the 
scope of the new type (which I dubbed "idx_t", by analogy from Idx in 
the regex code) to dfa.c. If we want to make this type visible we have 
more work to do, including publishing its maximum value (IDX_MAX in 
dfa.c), and making it easy to printf (I used "%td" in dfa.c, which is 
quite convenient but it assumes idx_t is ptrdiff_t).

From my use of idx_t in dfa.c I'm not entirely sold on the distinction 
between it and ptrdiff_t. The type is not separate from ptrdiff_t; it 
means "ptrdiff_t but nonnegative" and the code relies on that. There 
were iffy places where it kinda doesn't matter whether one uses 
ptrdiff_t or idx_t and where I expect other developers won't bother 
about the difference (as no tools check for it).

0001-dfa-prefer-ptrdiff_t-for-API-too.patch
Description: Text Data