Do we support any means of creating disjoint user-defined types, such as in SRFI-9, SRFI-99 or the R6RS record system?
I think providing a syntactic interface as the basis for records is more widely supportable and may offer better optimization chances, however the SRFI-99 API is too difficult to remember and RecordsArcfide seems only partially better, so I prefer the de facto standard SRFI-9 for WG1.
Several SRFIs, R6RS, and most Scheme implementations support some sort of uniform packed integer vectors. In particular, these are necessary for efficient binary I/O, and for memory mapping, so WG2 will certainly want them.
Do we provide a syntax and basic API for these in WG1?
I like the idea behind snellpym but it needs work - I'd rather give the author time to flesh it out. Failing that, I don't see what differentiates BlobAPI from the R6RS bytevectors.
R5RS provides a simple mechanism for easy cases of lazy evaluation. It does not support generalized lazy evaluation, because all built-in procedures are eager whether they 'need' to be or not. The relevant identifiers are delay and force; they are not present in IEEE Scheme. SRFI 45 argues that this pair is insufficient for expressing common lazy algorithms in a way that preserves tail recursion, and adds lazy (equivalent to (delay (force ...)), but acting atomically) and eager. The semantics of delay and force remain downward compatible.
Vote srfi-45 to add just the bindings lazy and eager in addition to delay and force, not all of the srfi-45 utilities. Vote none to remove delay and force from the standard.
Random numbers are useful for a wide variety of applications, including cryptography, statistics, games, simulations and genetic programming. Do we want to provide an interface to random number generation in WG1 Scheme?
Currently, there is no standard way to communicate with the context from which a Scheme program was started. This has become pretty standardized over time: a list of strings ("command-line arguments") and a map from strings to strings ("environment variables") on input, and a small integer or string on output ("exit value"). Scheme should recognize these realities.
We have command-line and exit from ModulesShinn, so the question remains if we should add SRFI-98 environment accessors.
In R5RS, many procedures and syntax forms return an "undefined value". In R6RS, the corresponding convention is to return "undefined values", meaning an undefined number (including zero) of undefined values. How shall R7RS go?
Vote r5rs for a single undefined value, r6rs for zero or more undefined values, or zero for exactly zero values. Anything other than r5rs would break R5RS (and IEEE) compatibility.
Too many existing programs expect exactly one value.
Assuming a single "undefined" value (dependent on the result of #68), users sometimes want to test for this value. If we enforce a unique undefined value, one approach is to generate this value explicitly to test against, such as (void) provided by some implementations. Alternately we could provide a test such as undefined?. Either approach breaks compatibility with existing extensions, and may be prohibitively difficult to implement for compilers lacking a separate undefined value type. Some programmers also consider testing for this value sign of a broken design.
Vote generate for a (void) procedure, test for undefined?, and both for both.
It's a bug to write programs which rely on this - unspecified is unspecified, and may be anything or even vary per compiler and program and call.
list?, length, equal? and other fundamental primitives may diverge when given cyclic data. In the former two cases, avoiding this is simple and not inefficient, and the equivalents are already provided in SRFI-1. In the latter case a proposal was made and rejected on the R6RS list. In the former case, R6RS seems to require list? return #f and length raise an error.
Do we want to specify the behavior when these primitives encounter cyclic data?
Options are equal? to specify equal? must terminate on cyclic input, r6rs to specify R6RS behavior for list? and length, srfi-1 to specify the SRFI-1 semantics (where length returns #f) and equal?+r6rs or equal?+srfi-1 are options for both.
equal? is dangerous to use if it may diverge. It would be reasonable to leave this unspecified, but since we already require shared structures checks for write it seems consistent to make the same requirement for equal?.
Old-fashioned Lisps used dynamic extent of variables. Although Scheme has switched to lexical scope, the concept of a dynamic environment can be useful in special cases.
Instead of special variables, SRFI-39 provides first-class "parameter" objects with dynamic bindings. Do we want to provide something similar?
Explicit mutation is the uncommon case, and I think it's safe to leave the semantics of this unspecified in the presence of threads. It's crucial, however, that parameterize be thread-local.
Short of a full time and date library, a single procedure
(current-seconds)
returning the epoch time in seconds, possibly as a real number, would be useful and is sufficient to implement a full library (though access to the host system's timezone would be desirable in that case).
Since some systems may not have access to a clock, we could make this an optional procedure. Alternately, it could be defined as a simple counter in such cases, providing an accurate notion of time ordering but no sense of duration. Finally, it could return #f in the absense of a clock.
This is a *huge* can of worms. POSIX time is simply a bug, and I would rather not have any time API than put it in WG1, but I want time to consider alternatives.
Should we have functions allowing a program to compute elapsed time, as distinct from calendar time?
TimeCowan contains a proposal.
I don't see the point of this. Time should be handled properly by #70 or not at all.
Should we provide case-lambda as in SRFI 16 and R6RS? It provides simple overloading of procedures based on the number of their arguments, and does not require that optional arguments appear only after mandatory ones.
case-lambda is widely implemented but I think encourages bad style. It's a terrible API when you want optional arguments, and is less expressive than a more general match-lambda.
It's not clear whether R5RS requires a PORT? procedure or not. It's listed in Section 3.3.2 under Disjointness of Types, but not under section 6.6.1 under Ports. R6RS requires it. Racket, Gauche, MIT Scheme, Gambit, Chicken, Guile, SISC support it; Scheme48/scsh, Kawa, and Chibi currently do not.
Shall we require it?
This is trivial to define and often provided.
Currently there's no way to determine whether a port is open or closed, short of trying to read/write to it and catching an error. Do we want to add an interface to this?
In most programs you should know the lifetime of the port, but in some cases this is necessary. I prefer port-open? because it's more common to want to test the "positive" capability of reading/writing before doing so.
In R5RS and R6RS, call-with-values takes two arguments, both procedures. The first is a producer of multiple values; the second is a consumer, to which the multiple values returned by producer are passed as arguments.
A possible extension is to allow multiple producer arguments, flattening all the produced values together, analogous to Common Lisp's multiple-value-call.
Do we add this extension?
Just because CL has something doesn't mean we should, and I haven't seen any convincing cases where this extension is useful.
SRFI-87 extends case with a => clauses, analogous to the use of => in cond clauses, which allows you to pass the item actually matched to a procedure.
Do we add this extension?
Syntactic sugar, rarely needed and easy enough to get around with an extra let.
SRFI-61 extends => clauses in cond with an optional guard form, such that after the value is generated and found to be true, it's further checked against the guard. If the guard returns #f the clause fails and processing proceeds to the next clause, otherwise the clause is accepted as normal.
Do we add this extension?
cond is complicated enough as it is.
Currently, => clauses in cond accept a single value from the generator (right-hand side) and pass it to the receiver (left-hand side). Shall we allow the generator to return multiple values and pass them to the receiver? If both this ticket and #89 pass, multiple values would also be allowed for generator/guard cond clauses.
Emphatically no. The whole point of => clauses is they are testing that the value generated is true - if multiple values are generated, which do we test for truth? Any semantics will be unintuitive to some. In addition, this will make => clauses slower in most implementations even if MV aren't used, because you need to account for them and box the result in the general case.
Should we allow (include "''filename''") at the REPL? This is distinct from import in that it just loads code without altering the module structure.
I don't see the point of this over load. The original ticket refers to load handling binary files whereas include would not, but in the proposed standard load isn't going to handle binary files anyway (beyond implementation-specific extensions).
The default reader in R7RS will default to case-sensitive, but users may wish to override this in some situations. R6RS allows at the top-level #!case-fold and #!no-case-fold read syntax to control the case-sensitivity of the current file. Many existing R5RS implementations, on the other hand, use #ci and #cs, with the difference that they refer to the next datum only.
Note PortsCowan provides a separate means of controlling case-sensitivity per-port.
Vote per-datum for the next-datum-only #ci/#cs syntax.
Per-datum flags can be handled entirely by read without any need for mutable state. I also dislike all #! forms (of which there are currently none).
The standard currently says nothing about the character encoding system of source files. Do we require this to be a fixed encoding such as UTF-8, use an out-of-band specification like the Emacs (and Python) -*- coding: foo -*- convention, or just leave it unspecified?
The emacs approach is handy but too much of a kludge to go into the small Scheme standard, and I don't want to force utf-8.
R6RS relegated string-set! to a module, and many modern languages tend towards making strings immutable. Removing entirely, however, breaks IEEE Scheme compatibility and should only be considered if you believe mutable strings are fundamentally broken.
Do we remove string-set!? Vote yes to remove, module to relegate to a module as in R6RS, or no to keep as is.
I consider mutable strings a design mistake in Scheme, but we need to preserve backwards compatibility so I prefer to discourage their mutation by putting them in a module. This is not just a symbolic gesture (like putting pair mutators in a module), because there are already existing Scheme implementations for which string-set! is very expensive.
In R6RS auxiliary keywords (such as else in cond and case forms) are explicitly exported from the (rnrs base (6)) library. Do we want to bind and export these from the core library?
If else is bound in the default module, then it must be imported at the call site whenever using it in cond or it won't match hygienically.
If else is not bound in the default module, then it must not be bound or imported at the call site whenever using it in cond or it won't match hygienically.
Another option is to specify for cond and case that they match the else identifier literally, ignoring any hygiene. This breaks compatibility with R5RS and R6RS.
else is conceptually unbound in the standard env and so should be specified as such (although for case it actually makes sense to match unhygienically since no other identifier would be legal).
In R5RS eqv?/equal? are in some sense the broadest tests for equality, comparing structural equality, but also tests for the same exactness, so that
(equal? 0 0.0) => #f
whereas
(= 0 0.0) => #t
Some users consider this confusing, others sometimes want an equal? that behaves like = for numbers.
Do we want to change equal? and eqv? in this way, or add a separate exactness-agnostic procedure? Vote yes to change, equal=? or inexact-equal? for separate procedures of those names (plus the analogous eqv=? or inexact-eqv?), or no to leave as is. Alternately, write in a separate name.
These are a common source of confusion, but I don't like breaking backwards compatibility, and think yet another function may prove even more confusing.
A bikeshed color issue, we need to choose the actual names for the module syntax for the winner of #2.
import, export and include are fairly universal and no alternate is suggested unless someone wants to write in a proposal.
The enclosing syntax can be called either library as in R6RS, module or some other proposal.
Similar to #102, we need to choose a name for the form to include Scheme code directly in a module form. This can be body as in the proposal, begin or some other name.
begin is used in the Scheme48 syntax, but it really doesn't mean the same thing as normal begin, and takes up an important binding in the module description language.
The include module form includes files literally with the default case-sensitivity. An include-ci form could include files case-insensitively without resorting to the reader hacks proposed in #92, allowing existing R5RS libraries to be used without modification.
This is easy to implement and is a nice way of providing backwards compatibility without any reader hacks.
Users invariably want some way to conditionally select code depending on the implementation and/or feature set available. CondExpandCowan allows conditional expansion in the style of SRFI-0 within the module language. SRFI-0 provides cond-expand, SRFI-103 provides a library naming extension, and numerous other personal hacks exist.
Do we want to include something along these lines in WG1 Scheme?
Something like this is very much needed. The search path extension is a hack, so I choose cond-expand, but I don't think a standard set of features belongs in WG1, even if they are optional.
R5RS specifies literal data in source code as immutable, but otherwise provides no way to generate or introspect immutable data.
One proposal is given in ImmutableData, providing mutable?, make-immutable and immutable->mutable.
Racket, for which all pairs are immutable in the default language, needs some way to generate shared and cyclic data structures at runtime, and provides the shared syntax for this. It also has an immutable? utility as the complement to mutable? above.
This is a complex topic which hasn't generated much discussion, and is better left to WG2. The ImmutableData proposal in particular is underspecified, and needs some discussion of whether it's a shallow or deep copy, how it handles closed variables, inherently mutable data-structures like ports, etc.
Currently equal? is strictly broader than eqv? except in the pathological case of comparing the same circular list with itself, for which eqv? returns true and equal? may loop infinitely. We could explicitly require equal? to check and return #t in this case, which most implementations do as a performance hack anyway.
This is an ugly special case to patch up something that should be handled by #51. If you're passing potentially circular structures to equal? at all in the absense of proper circularity handling you have a bug in your program.
R6RS provided a detailed exception system with support for raising and catching exceptions, using a hierarchy of exception types.
Do we use this, or parts of it, or a new exception system? The r6rs option is just for the core exception handling.
Exception systems have subtle semantics and we should not specify anything that hasn't even been implemented.
R5RS defines many things as "is an error" without any specification of what happens in that situation. R6RS goes to the opposite extreme and specifies as much as possible what exceptions are raised when.
Taking into account the system provided by ticket #18, we need to come up with guidelines for when exceptions should be raised, and clarify which R5RS "error" situations should raise an exception or be left unspecified.
R5RS specifies only 3 situations where an error is required to be signalled, leaving most situations unspecified as described in ErrorSituations.
I think there's probably a good line between r5rs and r6rs here, but no one has drawn it yet, and it's reasonable to stick with the r5rs default.
Do we provide any binary input or output ports, and if so how do we construct them and operate on them? Can binary and textual operations be mixed on the different port types?
PortsCowan provides binary port operations along with other extensions.
R6RS provides an entirely new I/O system, as well as a separate R5RS-compatible I/O system.
The withdrawn SRFI-91 provides yet another I/O system supporting binary ports.
Note this item as well as #29 and #31 specify semi-orthogonal aspects of I/O systems which are typically specified together by individual proposals. If the same proposal doesn't win for all three, the aspects will be merged as needed.
I think it's a mistake to _require_ implementations allow mixing of binary and character data, even if some implementations already do so.
Do we support encoding and decoding text from ports with different character encoding systems? Different end-of-line conventions? Different normalizations? How are encoding errors handled?
Do we provide a mechanism for custom ports, on which for instance string ports could be created?
R6RS as well as a number of Scheme implementations provide custom ports with various APIs.
R6RS and SRFI-69 both provide hash-table interfaces. Do we provide either of these, or try to provide some primitives on which efficient hash-tables can be implemented?
Exposing the eq?-hash function in SRFI-69 is a mistake.
We've decided to add file-exists? and delete-file, essential for a large class of scripts, but still have no way to get a list of files in a directory. Do we want to provide an interface to this?
let-syntax and letrec-syntax has known ambiguities in their behavior. We have the option of altering the semantics to correct this behavior, defining which behavior we intend, or removing let-syntax entirely. We could also leave this ambiguity unspecified.
The question of whether or not to introduce a new lexical scope (i.e. whether internal defines are visible outside the let-syntax) is straightforward.
If we don't introduce a new lexical scope, the question arises whether or not internal define-syntax forms are allowed and whether they apply to the body of the let-syntax, forms following the let-syntax, or both.
If internal define-syntax applies to the body, we may also wish to specify what happens when said define-syntax redefines an identifier bound by the enclosing let-syntax. This varies by implementation and may be difficult for macro expanders to change, so is left unspecified in the proposals below.
... and with the result of #6 also _ have special meaning in syntax-rules patterns, so they are not treated as pattern variables by default.
However their behavior when used in the literals list of syntax-rules is ambiguous, and simply breaks in most implementations.
Rather than breaking, it makes sense to go ahead and treat them as normal literals, overriding their special meanings.
In particular, there are many existing R5RS macros which make use of _ in the literals and are thus broken outright by #6. Allowing them as literals fixes these macros.
This fixes the R5RS macros that the new _ pattern breaks.
We need a naming convention for the core modules and standard libraries of the new module system.
In R5RS everything is effectively in a single module. R6RS provides a much more fine-grained breakdown of modules which could be retro-fitted to the bindings we include in our standard.
John Cowan has proposed a number of module factorings in items #71, #72, #73, #74, #75, #76, #77, as well as an I/O module breakdown in PortsCowan.
Since the naming and breakdown must be internally consistent I'm grouping these into a single ballot item. Members desiring to put forth a new proposal should specify where all bindings belong, or specify a subset of the bindings and default the rest to some other proposal.
Note some ballots specify explicitly whether or not the bindings in question are intended to be in a module or the core language. In these cases we still need to decide to which module they belong. Where specific votes contradict general factoring proposals, the specific vote wins.
Often a rational-only exponentiation function is useful; that is, a rational number raised to an integer power. Should we add this procedure to the core so that exponentiation is available even if inexact rationals are not provided or not imported?
I completely fail to see the point of this - is it motivated by some sort of efficiency concern, or perceived module factoring? Either rationale is misguided - expt is fine for rational as well as all other numbers.
NumericTower lists a plausible set of ten from fixnums only to the full monty. Which ones should we allow an implementation to provide? R5RS requires only fixnums large enough to handle string and vector indexes, while R6RS requires the full numeric tower.
Vote on the minimum level of support you want to require (implementations may of course still provide more than this). I've included only the most likely options below, write in other options if needed.
Note quaternions are a fairly rare numeric type, known to be provided only by extensions to scm and chicken, and thus may be difficult for other implementations to support if required.
We have a rich numeric tower, but there's no need to require the whole thing and rule out simple implementations.
R5RS provides quotient, modulo and remainder for integral division. R6RS extended this with div/mod and div0/mod0. A thorough analysis of possible division operations is provided in DivisionRiastradh, which includes a proposal for five separate division operator pairs. We need to choose which API we'll provide.
The R6RS operations are clearly bad, but I'm unconvinced we need everything provided by DivisionRiastradh.
In R5RS, symbols parsed as any sequence of valid symbol characters that does not begin with a character that can begin a number. The three exceptions +, - and ... are also provided. This allows parsing with only constant lookahead to determine type.
R6RS added additional exceptions for symbols beginning with ->, a common idiom, still allowing parsers to determine type with a constant lookahead.
John Cowan proposes allowing anything that cannot be parsed as a number to be a valid symbol. This removes the special case exceptions, but may require arbitrary lookahead.
Alex Shinn proposes symbols are any sequence of valid symbol characters that does not have a prefix which is a valid number. This removes the special case exceptions, allows constant lookahead, and allows extensions to number syntax.
It's important to leave room for numeric extensions, such as quaternions and units which are already provided by some Schemes. The cowan proposal makes this impossible. Ease of parsing, both by computers (not requiring arbitrary lookahead) and by humans (being able to tell if something is a number or symbol at a quick glance) is also a concern which should not be dismissed lightly. Both of these issues are addressed by the shinn proposal, which has just as simple a description and removes the special cases of R5RS and R6RS.
The WG has voted to have a list of character names.
The list in R5RS and the longer list in R6RS are only informative. I suggest adopting the R6RS list and making it normative.
The list is somewhat arbitrary, and I'd be open to other suggestions. Mostly I consider it important to make characters used for terminal manipulation avaiable, and to kill the vertical tab. The only character I debated on and ultimately left out was formfeed.
Similar to #84, we need to choose a specific list of mnemonic escapes like \n and \t to be recognized in strings.
These should be consitent with #84.