нашел старое из R рассылки
anonymous
August 10 2010, 11:54:20 UTC
On Mon, 8 Dec 2008, Stavros Macrakis wrote:
> I've read in many places that R semantics are based on Scheme semantics. As > a long-time Lisp user and implementor, I've tried to make this more precise, > and this is what I've found so far. I've excluded trivial things that > aren't basic semantic issues: support for arbitrary-precision integers; > subscripting; general style; etc. I would appreciate corrections or > additions from more experienced users of R -- I'm sure that some of the > points below simply reflect my ignorance. > > ==Similarities to Scheme== > > R has first-class function closures. (i.e. correctly supports upward and > downward funarg). > > R has a single namespace for functions and variables (Lisp-1). > > ==Important dissimilarities to Scheme (as opposed to other Lisps)== > > R is not properly tail-recursive.
True at present. May be unavoidable since the language provides access to the stack via things like sys.parent, but as it is rare to look at anything other than the immediate calling environment and call (outside of a debugging context) it may be possible to change that.
> > R does not have continuations or call-with-current-continuation or other > mechanisms for implementing coroutines, general iterators, and the like. > > R supports keyword arguments. > > ==Similarities to Lisp and other dynamic languages, including Scheme== > > R is runtime-typed and garbage-collected. > > R supports nested read-eval-print loops for debugging etc. > > R expressions are represented as user-manipulable data structures. > > ==Dissimilarities to all (modern) Lisps, including Scheme== > > R has call-by-need, not call-by-object-value. > > R does not have macros.
Those are related -- because of lazy evaluation one does macros are not needed to achive semantic goals (see for example tryCatch). Being able to define friendlier syntax would sometimes be nice though (see tryCatch again).
> R objects are values, not pointers, so a<-1:10; b<-a; b[1]<-999; a[1] => > 999. Similarly, functions cannot modify the contents of their arguments. > > There is no equivalent to set-car!/rplaca (not even pairlists and > expressions). For example, r<-pairlist(1,2); r[[1]]<-r does not create a > circular list. And in general there doesn't seem to be substructure sharing > at the semantic level (though there may be in the implementation). > > R does not have multiple value return in the Lisp sense. > > R assignment creates a new local variable on first assignment, dynamically. > So static analysis is not enough to determine variable reference (R is not > referentially transparent). Example: ff <- function(a){if (a) x<-1; x} ; > x<-99; ff(T) -> 1; ff(F) -> 99.
Correct, and a fair nuisance for code analysis and compilation work. I'm not sure how much would break if R adopted the conventions in Python (or with Scheme's define as I recall) that referencing a not yet initialized local variable is an error.
I'm not sure I would label this as meaning R is not referentially transparent thoug -- that goes out the window with mutable bindings as also available in Scheme.
> In R, most data types (including numeric vectors) do not have a standard > external representation which can be read back in without evaluation.
The default print form is not readable in this sense but dput is available for this purpose.
> R coerces logicals to numbers and numbers to strings. Lisps are stricter > about automatic type conversion -- except that false a.k.a. NIL == () in > Lisps other than Scheme.
A more important difference may be that logicals can have three values -- TRUE, FALSE and NA.
> I've read in many places that R semantics are based on Scheme semantics. As
> a long-time Lisp user and implementor, I've tried to make this more precise,
> and this is what I've found so far. I've excluded trivial things that
> aren't basic semantic issues: support for arbitrary-precision integers;
> subscripting; general style; etc. I would appreciate corrections or
> additions from more experienced users of R -- I'm sure that some of the
> points below simply reflect my ignorance.
>
> ==Similarities to Scheme==
>
> R has first-class function closures. (i.e. correctly supports upward and
> downward funarg).
>
> R has a single namespace for functions and variables (Lisp-1).
>
> ==Important dissimilarities to Scheme (as opposed to other Lisps)==
>
> R is not properly tail-recursive.
True at present. May be unavoidable since the language provides
access to the stack via things like sys.parent, but as it is rare to
look at anything other than the immediate calling environment and call
(outside of a debugging context) it may be possible to change that.
>
> R does not have continuations or call-with-current-continuation or other
> mechanisms for implementing coroutines, general iterators, and the like.
>
> R supports keyword arguments.
>
> ==Similarities to Lisp and other dynamic languages, including Scheme==
>
> R is runtime-typed and garbage-collected.
>
> R supports nested read-eval-print loops for debugging etc.
>
> R expressions are represented as user-manipulable data structures.
>
> ==Dissimilarities to all (modern) Lisps, including Scheme==
>
> R has call-by-need, not call-by-object-value.
>
> R does not have macros.
Those are related -- because of lazy evaluation one does macros are
not needed to achive semantic goals (see for example tryCatch). Being
able to define friendlier syntax would sometimes be nice though (see
tryCatch again).
> R objects are values, not pointers, so a<-1:10; b<-a; b[1]<-999; a[1] =>
> 999. Similarly, functions cannot modify the contents of their arguments.
>
> There is no equivalent to set-car!/rplaca (not even pairlists and
> expressions). For example, r<-pairlist(1,2); r[[1]]<-r does not create a
> circular list. And in general there doesn't seem to be substructure sharing
> at the semantic level (though there may be in the implementation).
>
> R does not have multiple value return in the Lisp sense.
>
> R assignment creates a new local variable on first assignment, dynamically.
> So static analysis is not enough to determine variable reference (R is not
> referentially transparent). Example: ff <- function(a){if (a) x<-1; x} ;
> x<-99; ff(T) -> 1; ff(F) -> 99.
Correct, and a fair nuisance for code analysis and compilation work.
I'm not sure how much would break if R adopted the conventions in
Python (or with Scheme's define as I recall) that referencing a not
yet initialized local variable is an error.
I'm not sure I would label this as meaning R is not referentially
transparent thoug -- that goes out the window with mutable bindings as
also available in Scheme.
> In R, most data types (including numeric vectors) do not have a standard
> external representation which can be read back in without evaluation.
The default print form is not readable in this sense but dput is
available for this purpose.
> R coerces logicals to numbers and numbers to strings. Lisps are stricter
> about automatic type conversion -- except that false a.k.a. NIL == () in
> Lisps other than Scheme.
A more important difference may be that logicals can have three values
-- TRUE, FALSE and NA.
luke
Reply
Leave a comment