Document Number: WG14 N686/X3J11 __-___
C9X Revision Proposal
Title: Obsolescent Features of the C Standard_______________
Author: Peter Seebach_______________________________________
Author Affiliation: ________________________________________
Postal Address: 3949 Lyndale Avenue South, Minneapolis MN 55409
E-mail Address: firstname.lastname@example.org_____________________________
Telephone Number: +1 612-824-6045___________________________
Fax Number: ________________________________________________
Sponsor: Peter Seebach______________________________________
__ Editorial change/non-normative contribution
__ New feature
__ Addition to obsolescent feature list
__ Addition to Future Directions
Y_ Other (please specify) Removal of old features_______
Area of Standard Affected:
__ Macro/typedef/tag name
__ Other (please specify) ______________________________
Prior Art: Various, including GNU C_________________________
Target Audience: Anyone not still living in the 80's________
Related Documents (if any): ________________________________
Proposal Attached: Y_ Yes __ No, but what's your interest?
C89 declared several items to be "obsolescent features"; these
features were slated for removal in future versions of the C standard.
This paper proposes ways of removing them.
There are several separate obsolescent features to be looked at.
I will use the C9X draft 9 numbering, which is off by one from the C89
6.9.2: I don't show a vote taken on this, so I'm ignoring
6.9.5: The straw vote was 7 to implement or remove this, 5 to
leave it alone, and 5 abstentions. There are at least three ways
to handle this:
1. Remove paragraph 6.9.5. This is the simplest.
2. Implement this restriction, by adding wording rendering
the use of two array-declared parameters for the same
3. In 6.7.1, paragraph 9, replace
shall be adjusted to ``pointer to type,''
shall be adjusted to ``restrict-qualified pointer
so that the rule is enforced for us.
I have heard that the Borland C compiler already implements something
equivalent to the 2nd or 3rd suggestion. (It also doesn't realize
that the adjustment to pointer type implies that a parameter declared
as an array is actually a pointer which can be modified; rather, it
treats them as normal arrays.)
The third provides a convenient way for compilers which are using
restrict for optimization purposes to be compatible with code that
may need to run on older machines. Any code which is affected by
this is, as far as I can tell, already covered by the obsolescent
feature list, and any code using this idiom would be guaranteed
to work as expected on C89 compilers.
6.9.3 and 6.9.4:
These are really very deeply tied. I've gotten a lot of feedback from
the user community, and the general impression I've gotten is that
old-style definitions are very popular, but no one will miss old-style
declarations. The old-style definitions are popular because they
allow for convenient grouping of parameters of like type, and make it
easier to break function definitions up into easily parsed lines,
while keeping the names of arguments all together and in order.
Removing all compatibility with old-style declarations and definitions
is not very hard. It will break a sizeable body of existing code, but
there has been plenty of warning, and real compilers will not stop
having 'traditional' modes in this decade.
Another alternative, and one for which there is prior art, is to
suppress the usual argument promotions when a prototype for the
function is already in scope. This does not change the behavior of
K&R code at all, and affects the behavior of no C89 program which does
not invoke undefined behavior.
Yet another is to simply eliminate the usual argument promotions
entirely, and allow the old-style definitions as an alternative
way to format a prototyped definition.
The first option is probably the simplest, the second is probably the
most convenient for users, and is the only one for which I have seen
any existing practice, and the third is probably the most internally
consistent but still convenient for users.
NetBSD, and also at least one of the other *BSD's, has a ~150 megabyte
source tree which is currently using option 2; all functions are
prototyped, but old-style definitions are used for almost everything.
This appears to be compatible with K&R systems, and also usable for
the GNU C compiler.
The language extension used appears to suppress the default promotions
only when a prototype is in scope.
To implement option 1, we would remove the paragraphs describing the
obsolescent feature, and simply remove the majority of the text
describing interactions between identifier-list and
parameter-type-list function declarations. To implement option 3, we
would remove all of the rules for the default promotions. To
implement option two, we would have to add some clever wording in
the paragraph about identifier-list vs. parameter-type-list function