quiroz@cs.rochester.edu (Cesar Quiroz) (01/13/88)
Both William Sommerfeld and Chris Torek propose handling compiler
hints as assertions (not necessarily the assert macro, of course).
As this was still an idea in the making, they offered no concrete
syntax for it, as is reasonable at this stage. Thus, it would be
unfair to criticize those ideas on syntactic grounds. Nevertheless,
both proposals showed `freehand' examples of how it could work, and
both examples were of the type I call "pointwise" in the discussion
below. Just to make sure a point I made before doesn't get lost in
the noise, let me insist that such hints are likely to need some
indication of scope too.
The models for hinting a C compiler (William Sommerfeld rightly
notices that this belongs on the theory of hinting *any* compiler)
that are in discussion are like this:
`Pointwise': `Scoped':
Assertions tell the compiler Assertions, as in the pointwise
something it might not be able model, plus (optional?) scope of
to figure out by itself. applicability.
Ex: for (...;....;...) { Ex: for (...;....;...) {
...code... ...code...
assertion,... assertion {
...code... .. code where assertion is
} .. said to be valid
}
It is left to the compiler to ... code where assertion is
propagate the information given, ... not guaranteed to be valid
so it will presumably have to decide }
when it ceases to be valid.
The reason for recommending the scoping is precisely that these hints
are useful when the compiler doesn't know how to figure them out by
itself. It could hardly be expected to propagate this information,
although it might be possible in some cases.
For instance, when disambiguating array references by guaranteeing
that two integer expressions are not equal, we are sparing the
compiler having to solve Diophantine equations. So we can assume
the compiler doesn't know how to solve such, how could it know if
the given information holds in the next line? (Add adequate amounts
of handwaving until this sounds likely).
Chris Torek's "clean wound" principle sounds very sensible, but I
suspect C is inhospitable to such extensions. The notation needed
to express the class of hints that are interesting would be a
programming language of higher level. Perhaps a good subject for
comp.lang.misc.
(To think that all this is just warm up for the day when we really
know what `noalias' looks like :-)
--
Cesar Augusto Quiroz Gonzalez
Department of Computer Science ...allegra!rochester!quiroz
University of Rochester or
Rochester, NY 14627 quiroz@cs.rochester.edu