Path: utzoo!mnetor!uunet!seismo!sundc!pitstop!sun!decwrl!decvax!mcnc!gatech!udel!princeton!phoenix!rjchen
From: rjchen@phoenix.Princeton.EDU (Raymond Juimong Chen)
Newsgroups: comp.lang.c
Subject: Another 'D' proposal (sigh)
Message-ID: <1893@phoenix.Princeton.EDU>
Date: 28 Feb 88 20:05:15 GMT
Organization: Princeton University, NJ
Lines: 67

I don't have a copy of dpANS so I don't know if this is addressed
in it.  I also don't know if this has been discussed and rejected
by the Committee, so here goes...

Propose that there be a way to declare "true functions", ie, functions
whose return values depend on and are solely determined by the arguments
passed to it.  They also have no side-effects.

Simple examples:  min() max() abs() pow() time_diff() etc.

This would be a hint to the compiler that such function calls have
no side-effects and therefore can

	1. be optimized away in expressions such as
		if (0 * min(a,b)) ...

	2. be calculated only once if the arguments are identical
	   in different statements:
		foo = (time_diff(start, end) + bar) * snerg;
		/* ... statements which do not alter start or end */
		oog = time_diff(start, end);

	3. be calculated at any time provided the arguments are known.

This also allows you to use such "true functions" in DEBUG macros
without having the compiler generate idle code.  for example,
	DEBUG(("x^2 is now %f\n", pow(x,2)));
(Normally, the compiler would have to generate a call to pow(), since
it doesn't know that pow() has no side-effects--even when DEBUG if
#define'd to a null string.)

Implementation:
I can think of two ways this could be implemented.  Let foo() be a
"true function".

register int foo(...)
const int foo(...)

The first method is consistent with the notion of "register" as a
"hint to the compiler".  The second is consistent with the notion of
"const" as "unchanging".  (Observe that the use of "const" does not
conflict with soon-to-be-existing usage.
	const char * foo(...)
and	char * const foo(...)
would mean different things.  One says that foo() returns a pointer to
a character in ROM, the other says that foo() returns a pointer to a 
character which depends purely on the arguments.  I don't know which
is which, however.)

The compiler can even check whether you're lying or not:  If inside
your function you refer to anything other than an argument or a
local variable, or if you call a "non-true" function, then
the "trueness" of your function is suspect and can be
flagged by a suitably picky compiler (or lint).

Comments welcome.
--
Raymond Chen	UUCP: ...allegra!princeton!{phoenix|pucc}!rjchen
		BITNET: rjchen@phoenix.UUCP, rjchen@pucc
		ARPA: rjchen@phoenix.PRINCETON.EDU
"Say something, please!  ('Yes' would be best.)" - The Doctor

-- 
Raymond Chen	UUCP: ...allegra!princeton!{phoenix|pucc}!rjchen
		BITNET: rjchen@phoenix.UUCP, rjchen@pucc
		ARPA: rjchen@phoenix.PRINCETON.EDU
"Say something, please!  ('Yes' would be best.)" - The Doctor