#pragma

[Chris Torek]

In article <12570056 at hprsue.HP.COM> sue at hprsue.HP.COM (Sue Meloy) writes:
>The only safe thing is for all implementation-specific directives 
>to be protected by ifdefs.

(sounds good to me)

>For example (no noalias diatribes, please!),

Why not? :-)

>do_something(int *vector1, int *vector2)
>#ifdef WHIZBANG_C
>#pragma noalias vector1,vector2
>#endif

I still think `noalias' is a horrible name.  For those who want to
build compilers without automatic alias detection (which is, I admit,
both impossible in general and very slow in the usual possible cases),
I suggest the syntax

	noninterfering(p1, p2, p3, ..., pn)

which means that stores through p1 do not `interfere' with indirection
through any of p2, p3, ..., pn; stores through p2 do not interfere with
indirection through p1, p3, ..., pn; ...; and stores through pn do not
intefere with indirection through p1, p2, ..., pn.

Note that this is a weaker restriction.  It should probably be augmented
to take functions as well:

	noninterfering(p1, munch)

would say that munch() never affects p1[i] for any legal i (and, equivalently,
that p1[i] never affects munch()).

With a special syntax (ick) this same declaration could declare a pointer
or function `pure':

	noninterfering(sin, *)	# sin(x) is a pure function
	noninterfering(ptr, *)	# ptr points to memory not affected by
				# any other pointer or function

It is not clear how to indicate that second-level pointers are pure.
Something along the lines of

	double **x;
	#pragma noninterfering(x, x[*], *)

might declare that not only is x[i] `pure', but also x[i][j].

This syntax could almost certainly be improved.

Oh well, enough of this.
-- 
In-Real-Life: Chris Torek, Univ of MD Comp Sci Dept (+1 301 405 2750)
Domain:	chris at cs.umd.edu	Path:	uunet!mimsy!chris