https://stackoverflow.com/questions/16651843
italiano
english
français
española
中国
日本の
العربية
Deutsch
한국어
Português
RussianCommon Lisp has a detailed file compilation model as part of the standard language. The model supports compiling the program to object files in one environment, and loading them into an image in another environment. There is nothing comparable in Scheme. No eval-when, or compile-file, load-time-value or concepts like what is an externalizable object, how semantics in compiled code must agree with interpreted code. Lisp has a way to have functions inlined or not to have them inlined, and so basically you control with great precision what happens when a compiled module is re-loaded.
By contrast, until a recent revision of the Scheme report, the Scheme language was completely silent on the topic of how a Scheme program is broken into multiple files. No functions or macros were provided for this. Look at R5RS, under 6.6.4 System Interface. All that you have there is a very loosely defined load function:
optional procedure: (load filename)
Filename should be a string naming an existing file containing Scheme source code. The load procedure reads expressions and definitions from the file and evaluates them sequentially. It is unspecified whether the results of the expressions are printed. The load procedure does not affect the values returned by current-input-port and current-output-port. Load returns an unspecified value.
Rationale: For portability, load must operate on source files. Its operation on other kinds of files necessarily varies among implementations.
So if that is the extent of your vision about how applications are built from modules, and all details beyond that are left to implementors to work out, of course the sky is the limit regarding inventing programming language semantics. Note in part the Rationale part: if load is defined as operating on source files (with all else being a bonus courtesy of the implementors) then it is nothing more than a textual inclusion mechanism like #include in the C language, and so the Scheme application is really just one body of text that is physically spread into multiple text files pulled together by load.
If you're thinking about adding any feature to Common Lisp, you have to think about how it fits into its detailed dynamic loading and compilation model, while preserving the good performance that users expect.
If the feature you're thinking of requires global, whole-program optimization (whereby the system needs to see the structural source code of everything) in order that users' programs not run poorly (and in particular programs which don't use that feature) then it won't really fly.
Specifically with regard to the semantics of continuations, there are issues. In the usual semantics of a block scope, once we leave a scope and perform cleanup, that is gone; we cannot go back to that scope in time and resume the computation. Common Lisp is ordinary in that way. We have the unwind-protect construct which performs unconditional cleanup actions when a scope terminates. This is the basis for features like with-open-file which provides an open file handle object to a block scope and ensures that this is closed no matter how the block scope terminates. If a continuation escapes from that scope, that continuation no longer has a valid file. We cannot simply not close the file when we leave the scope because there is no assurance that the continuation will ever be used; that is to say, we have to assume that the scope is in fact being abandoned forever and clean up the resource in a timely way. The band-aid solution for this kind of problem is dynamic-wind, which lets us add handlers on entry and exit to a block scope. Thus we can re-open the file when the block is restarted by a continuation. And not only re-open it, but actually position the stream at exactly the same position in the file and so on. If the stream was half way through decoding some UTF-8 character, we must put it into the same state. So if Lisp got continuations, either they would be broken by various with- constructs that perform cleanup (poor integration) or else those constructs would have to acquire much more hairy semantics.
There are alternatives to continuations. Some uses of continuations are non-essential. Essentially the same code organization can be obtained with closures or restarts. Also, there is a powerful language/operating-system construct that can compete with the continuation: namely, the thread. While continuations have aspects that are not modeled nicely by threads (and not to mention that they do not introduce deadlocks and race conditions into the code) they also have disadvantages compared to threads: like the lack of actual concurrency for utilization of multiple processors, or prioritization. Many problems expressible with continuations can be expressed with threads almost as easily. For instance, continuations let us write a recursive-descent parser which looks like a stream-like object which just returns progressive results as it parses. The code is actually a recursive descent parser and not a state machine which simulates one. Threads let us do the same thing: we can put the parser into a thread wrapped in an "active object", which has some "get next thing" method that pulls stuff from a queue. As the thread parsers, instead of returning a continuation, it just throws objects into a queue (and possibly blocks for some other thread to remove them). Continuation of execution is provided by resuming that thread; its thread context is the continuation. Not all threading models suffer from race conditions (as much); there is for instance cooperative threading, under which one thread runs at a time, and thread switches only potentially take place when a thread makes an explicit call into the threading kernel. Major Common Lisp implementations have had light-weight threads (typically called "processes") for decades, and have gradually moved toward more sophisticated threading with multiprocessing support. The support for threads lessens the need for continuations, and is a greater implementation priority because language run-times without thread support are at technological disadvantage: inability to take full advantage of the hardware resources.
