We voted that eqv? return #t if both arguments are any value which writes as +nan.0. The description of this item was ill-formed and confusing, as objected to in:
http://lists.scheme-reports.org/pipermail/scheme-reports/2011-September/001507.html
We therefore are re-opening the item, with amended descriptions.
The same proposal is what R6RS specifies - that we add a clause to the definition of eqv? saying that if both arguments are NaN values with the same bit pattern, eqv? must return #t. Thus eq? implies eqv?. However, if two values both print as +nan.0 they may or may not be eqv?. This also requires additional checks for floating point comparisons.
The different proposal is that we add a single clause requiring (eqv? +nan.0 x) to return #f for any x. This is the behavior that results for any R5RS implementation that adds support for +nan.0 as an IEEE float without any special handling for it in eqv?.
The unspecified proposal is to make the results explicitly unspecified.
The following 8 implementations return #t: Chez, Gambit, Guile, Ikarus/Vicare, Kawa, Larceny, Racket, STklos.
The following 6 implementations return #f: Bigloo, Chibi, Chicken, Gauche, MIT Scheme, Scheme48.
SigScheme and Scheme 9 don't have +nan.0. SISC currently has a bug where (= nan.0 x) is true for any x.
We need to handle this directly in the definition of eqv?. As the current definition stands, if specified at all it must be different.
Excluding -nan.0 was an oversight, and it's gratuitously incompatible with R6RS as well as current practice. Racket, Gauche, Chicken, Guile, Chez, Ikarus, Larceny, Ypsilon, STklos all support +nan.0 and -nan.0 as equivalent forms. MIT, Bigloo, Scheme48/scsh, SISC, SCM, Scheme 9 don't support either form. Only Gambit and Chibi support +nan.0 but not -nan.0.
STklos prints both +nan.0 and -nan.0 as -nan.0.
Vote yes to allow -nan.0, no to disallow it.
I don't understand the motivation behind this. NaN can neither be positive nor negative - the "+" is just a hint that the value is numeric (which is no longer true with the symbol extensions that now allow an initial "+"). Thus "-" serves no purpose and does not simplify the syntax - if we allow multiple forms, do we also allow "nan.0", "+nan.", "+nan", "+nan.1", etc.? One is easier to remember than two.
Bradley Lucier says:
I don't see the centered-* operators as somehow a "completion" of the other division operators. In the small language I'd recommend only the truncate-* and floor-* operators for two reasons: they are the only division operators that have an established history of use in computer programming and mathematics, and they form a minimal extension of R5RS. (I'm not saying that the other division operators have never been used in mathematics or programming (see CL), but small Scheme is not supposed to be a kitchen-sink language.)
Vote shrink to prune to truncate-* (R5RS) and floor-* (R5RS modulo), moving the extra operators to the large language, or keep to keep all 18 division operators in the small language.
I agree completely. The operators need more use and individual rationales before they are promoted to the small language. I'm also voting shrink/core to suggest we put them back in the core language, iff we shrink.
Currently vectors are the only type represented by a readable datum that are neither self-quoting nor meaningful Scheme expressions (i.e. symbols and lists). The proposal is to make them self-quoting as well.
Currently Racket, Gauche, MIT, Guile, Kawa, Chibi, SCM, STklos, Scheme 9, Scheme 7, UMB, VX, Oaklisp treat vectors as self-quoting.
Gambit, Chicken, Bigloo, Scheme48/scsh, SISC, Ikarus, Larceny, Ypsilon, IronScheme, Mosh, KSi, SigScheme, Elk treat unquoted vectors as errors.
Vote yes to make them self-quoting, no to make it an explicit error, or unspecified to leave unspecified as in R5RS.
This is fairly widely supported and is the only reasonable semantics for unquoted vectors.
The specifications of map, for-each, and other procedures that accept a procedure as an argument and call it, should specify that the argument procedures will always be called in the dynamic environment of the call to map, for-each, etc.
This is an R6RS fix.
Vote yes to add the clarification and no to leave it out.
That the semantics are desired is obvious, but I don't see how they could be interpreted otherwise and I'm not convinced it's worth writing this.
Identifiers beginning with a character of type Nd, Mc, or Me should be forbidden. This is an R6RS issue.
Nd is a numeric character, which in the case of ASCII 0-9 is already forbidden, but currently unspecified for non-ASCII digits.
Mc and Me are enclosing marks and spacing combining marks respectively, which are logically attached to the preceding character.
Vote yes to forbid (which would still allow this as an implementation-dependent extension for either numbers or symbols).
This makes sense - such identifiers would look like numbers which would be misleading.
This is equivalent to eq? on symbols, and provides R6RS base compatibility as well as completing the set of type-specific comparisons. See also #316.
Vote yes to add this procedure.
This helps with type-inference, although in practice the large language will want to provide some form of identifier=?.
Real-valued?, rational-valued?, and integer-valued? test whether a given number object can be coerced to the specified type without loss of numerical accuracy. They are equivalent to the versions of real?, rational?, and integer? that exist in R5RS.
Specifically, the behavior of these predicates differs from the behavior of real?, rational?, and integer? on complex number objects whose imaginary part is inexact zero.
These procedures provide R6RS base compatibility as well.
Vote yes to add these three procedures.
If nothing else the names are too confusing - the difference is too small, and I don't think people will be able to keep these straight.
Assert raises an error if its argument is #f. This provides R6RS base compatibility.
Vote basic to add this syntax. Vote optionals to make assert optionally accept, after its expression argument, a single message argument and zero or more irritant arguments in the same manner as the error procedure. Vote no in order not to add assert.
There are numerous assert macros, and it's not clear which is best - in particular, the best ones allow friendlier reporting. This isn't ready for the small language.
Infinite? returns #t if its value is a real number, or if its value is a complex number and either the real or the imaginary part would return #t to infinite?. This provides R6RS base compatibility, with extensions for complex numbers analogous to that provided by finite? and nan?.
This was in the draft at one point, but was never actually voted on, so the editors removed it.
Vote yes to add this procedure.
I more often want to check infinite? than finite?.
Bradley Lucier writes (lightly edited):
A square primitive is useful in calculating with bignums because squaring a bignum is generally cheaper than multiplying two different bignums of the same size. For example, Gambit's runtime checks trivially whether the two arguments in (* a b) are eq? before calling the appropriate algorithm. Generally, it may be better to be able to express this primitive directly.
[He also points out that given square in the small language, we can have flsquare in the large language, though having the latter doesn't actually require having the former.]
In addition, there are 20,340 Google hits for ["(define (square x)" ss|scm].
Vote yes to add this procedure.
On the one hand, (expt x 2) is the same thing, and should be just as easy to optimize. On the other, we provide sqrt as an alias for (expt x 1/2) and it seems odd not to have this inverse.
For with-input-from-file, with-output-to-file, call-with-input-file, call-with-output-file, R5RS just says that the file should exist. However, open-input-file requires an error to be signalled if the file cannot be opened, whether because it does not exist for some other reason like the lack of permissions. This inconsistency doesn't seem useful.
The proposal is to change these wrapper procedures to also require an error to be signalled if the file cannot be opened. All major Schemes already implement this.
Vote yes to require signalling an error if the files cannot be opened.
I think R5RS implies this behavior, but it's worth making explicit.
mdmkolbe writes on Slashdot:
Given that on a system with Unicode, you almost never want to do a non-normalizing case-insensitive match and that it is hard for a user to efficiently implement their own normalizing case-insensitive match, it seems an odd corner of the rectangle to omit.
(end quotation)
Alternatively we could specify that -ci procedures always normalize, or that -ni procedures are always case-insensitive, since the details of the normalization are not exposed anyway.
I think this point suggests that the normalization API doesn't compose well, and needs to be thought out better. I'd rather remove it and provide something more powerful in the large language. Otherwise, we should leave it as is.
Currently nothing is said about the <test> of if, cond, and, or, etc. returning zero values or multiple values. The proposal is to make this an explicit error. Remember that this does not mean an error is signalled.
Vote yes to make an explicit error.
As the standard already says, "Except for continuations created by the call-with-values procedure [...] all continuations take exactly one value. The effect of passing no value or more than one value to continuations that were not created by call-with-values is unspecified." Repeating this information for every continuation that could not be created by call-with-values is redundant.
Currently nothing is said about what happens if the <expression> of a set! returns zero values or multiple values. The proposal is to make this an explicit error. Remember that this does not mean an error is signalled.
Vote yes to make an explicit error.
Same argument as for #293 above.
Right now nothing is said. The proposal is to make this an explicit error. Remember that this does not mean an error is signalled.
Vote yes to make an explicit error.
Same argument as for #293 above.
The case-folding flags #!fold-case and #!no-fold-case are the only reader flags in the draft, however their need is reduced (though not eliminated) by the library declaration include-ci. Do we still need flipflop flags to turn case-folding on and off in part of a file?
If we remove these we maintain backwards compatibility with R5RS library code, however we lose the ability to support R5RS programs or toggle case-folding in the REPL or data files, etc.
We could only remove these in conjunction with a proposal that allows toggling case-folding in the REPL.
[Based on feedback from Marc Feeley.]
delay and force were simple balanced concepts, but the introduction of lazy somewhat confuses the issue - when is delay appropriate and when is lazy? A simple solution would be to rename lazy to delay-force, indicating it is simply the composition of delay and force, and letting people see directly in code the balance of delays and forces.
lazy is confusing, delay-force makes the usage and the relation to existing operators obvious.
[Based on feedback from Marc Feeley.]
Currently symbols can either be delimited with pipes |...| with optional hex escapes inside, or include hex escapes directly without the pipes. This wastes two characters that were reserved in R5RS, the pipe and the backslash, when either one by itself would be sufficient to represent all symbols. This is especially unfortunate because both characters are used as extensions in various Schemes - the pipe being another symbol character in SCSH (to represent shell-style pipes and C-style operators) and the backslash used in Gambit's infix syntax. We should reconsider if we really need to take up both of these characters.
Both is a waste - I prefer either-or. Although the uses in SCSH are nice, the |...| syntax is very widely implemented.
They have been required for a long time, but Alex Shinn says:
I definitely think everything but the one and two depth combinations should be removed from (scheme base). Their use is generally a code smell. People should use destructuring, records, or SRFI-1 first..tenth accessors.
Ray Dillinger (Bear) adds:
The historic use of these entities was as accessors for structured aggregates implemented with cons cells. In a language that directly supports records, they have a reduced mission.
Vote base to keep all in the base library or library to move the 3- and 4-letter accessors to a separate library.
I think removing them is too strong, but would like to be able to trim down the base library of 24 marginal-use procedures.
[Based on feedback from Marc Feeley]
The eager procedure is named particularly unfortunately because it sounds as though it is in some way paired with lazy, and there is anecdotal evidence it was voted in on this misunderstanding. In fact, it is completely unrelated to lazy, being just a utility procedure that has never been seen used in practice. Perhaps a better name for it would be promise or make-promise, since it just creates an (already computed) promise value.
Vote eager, promise or make-promise to specify the name, or remove to remove this procedure altogether.
I suspect people voted this in under the false impression it formed a pair of some sort with lazy. The utility is minor so I'd just as soon leave it out, but if we do keep it I'd prefer a self-explanatory name.
Allow the argument of list-ref to be circular. It is still an error to use an index >= the length of the list. None of my test implementations has a problem with this.
Vote circular to explicitly allow circular lists, error to add an "is an error" disclaimer, or unspecified to leave as is.
A de-facto standard - every implementation allows this anyway.
Allow circular lists as the list arguments to map and for-each. If all arguments are circular, these procedures will not terminate unless the mapping procedure forces a non-local exit. The result of map is not circular. Implementations that stop when the shortest list runs out and don't make gratuitous tests shouldn't have a problem with this: R5RS allows, R6RS forbids, and R7RS requires this behavior.
Vote circular to explicitly allow circular lists, error to add an "is an error" disclaimer, or unspecified to leave as is. Unspecified leaves open the theoretical extension of returning a new circular list with the corresponding mapped results.
The draft R7RS semantics makes this obvious.
When we approved CompleteSequenceCowan in ticket #64, we adopted SRFI 43 syntax and semantics for vector-copy, meaning that it takes optional start, end, fill arguments. This is inconsistent with various other copier procedures in R7RS as inherited from R5RS, as well as what is provided in SRFI 43 and its relatives SRFI 1 (for lists) and SRFI 13 (for strings). There are four plausible courses of action:
Simpler is better in the core - people can always competing SRFIs and utility libraries.
The current draft guarantees the guard clauses (not the body) of a guard form to be in tail call position, but the need for this is unclear (who needs an unbounded number of active exceptions), and there may be worthwhile guard implementations where this is not the case.
A minor detail, but I think mentioning detracts more than it adds from the standard.
The current draft allows @ to begin an identifier, which would require some comment about unquoting, i.e. to distinguish whether ,@foo is (unquote @foo) or (unquote-splicing foo).
The options are invalid (disallow @ at the beginning of an identifier, as in R5RS), unquote to indicate that ,@foo is (unquote @foo), and unquote-splicing to indicate that ,@foo is (unquote-splicing foo).
If unquote-splicing is chosen, a note will be added saying that if you want to unquote an identifier beginning with @ you need to either insert whitespace or escape the identifier, e.g. either , @foo or ,|@foo|.
Note that if we don't choose invalid then SXML retroactively becomes valid syntax.
This is what all implementations I'm aware of do - unquote would be strange and difficult semantics and is not completely speficied here, while invalid is undesirable due to SXML.
We should probably make (string-set! str n #\null) unspecified. Note that R7RS implementations can already restrict the set of characters that are allowed in strings.
Vote yes to add a clause to this effect, and no to leave it as legal.
Many implementations can represent the null char, and at the same time use C strings as the underlying string representation, which makes this unspecified.
This is equivalent to eq? on booleans, and provides R6RS base compatibility as well as completing the set of type-specific comparisons. See also #285.
Vote yes to add these three procedures.
As in #285, this helps with type-inference.
The draft states for with-input-from-file and with-output-to-file:
If an escape procedure is used to escape from the continuation of these procedures, their behavior is implementation-dependent.
but now that we have dynamic-wind there's no particular reason to keep this restriction, nor is it difficult to implement.
Vote parameterize to specify the current-in/output-port are bound dynamically as with parameterize in these cases, or unspecified to leave unspecified.
This was pretty clearly an oversight in R5RS which already had dynamic-wind.
Currently we require that if the characters GREEK LETTER CAPITAL SIGMA, SMALL SIGMA, and SMALL FINAL SIGMA are supported by an implementation, that a CAPITAL SIGMA in a string passed to string-downcase be changed to SMALL FINAL SIGMA just before a word break, and SMALL SIGMA otherwise. Word breaks are defined by UAX #29, and are no simple matter. The proposal is to make this behavior optional, allowing CAPITAL SIGMA to be downcased to SMALL SIGMA in every case.
Vote yes to make optional.
I think this needs more discussion, but if we're defining string-*case in terms of an AI-complete word-break operator then that's the real problem, not what to do with sigma.
(In this ticket, "complex" is used for readability; it is synonymous with "non-real".)
This feature is true in implementations that support complex numbers such that both the real and the imaginary parts are exact; that is, if (eqv? 3+4i 3.0+4.0i) evaluates to #f. This feature is false if complex numbers are not supported or if only inexact complex numbers are supported. Most of the applications of complex numbers use inexact numbers, but some applications may require exactness: this feature allows those applications to fail fast on implementations that cannot support them.
Existing implementations:
Vote yes to add this feature.
This is handy - it is already being used in the Chibi codebase.
This procedure returns a list of symbols corresponding to the feature identifiers which the implementation treats as true. More details at EnvironmentEnquiriesCowan.
Vote yes to add this procedure.
This can be useful for debugging and version info.
EnvironmentEnquiriesCowan is a library providing at run time what Common Lisp calls environment enquiries such as the name of the OS. Implementations can currently expose these as cond-expand feature identifiers, but there is no way to determine things like the name of the implementation at run time so that it can be written to a log file, for example.
Vote yes to add EnvironmentEnquiriesCowan (other than get-features), and no to leave out.
This needs more use and investigation, and as a library has no advantage being in the small language (as opposed to features). Let's add this to the large language.
Reduce the standardized cond-expand feature identifiers to r7rs, exact-closed, ratios, ieee-float, and full-unicode, plus the name and name-plus-version of the implementation. The others can't affect the behavior of strictly conforming programs, and it's not clear if they apply to compile time or run time on implementations that distinguish the two. See also ticket #320 for exact-complex.
Argument against: Keeping them in the standard encourages all implementations that use them to spell them the same way: darwin, not macosx.
Vote full to keep the full list as in draft-6, implementation to keep only the implementation features, or numerics to keep the list described above.
Pending a better list, I think many of the features are useful, and are actively in use in several implementations.
The (library <name>) feature test which is true if the given library is available (at compile time). This was used because we voted for CondExpandCowan, but the original syntax was just <name> which is ambiguous and therefore invalid. The switch to (library <name>) was added editorially, but not officially voted on.
Vote keep to keep and remove to remove.
A formality - we basically already voted this in.
Allow \| to represent a vertical bar in an identifier enclosed in vertical bars (the current BNF disallows | anywhere in the escape).
Note this item is nullified if |...| escapes are removed in item #304.
Vote pipe to allow just the vertical bar escaped, string to allow the same set of escapes as in string literals (plus pipe), and none to leave as is.
I think this was an oversight.
(bytevector-copy! from to) is equivalent to (bytevector-copy-partial! from 0 (bytevector-length) to 0).
The proposal is to remove the existing bytevector-copy! from the small language, and rename bytevector-copy-partial! to bytevector-copy!, with the order of arguments `to at from start end, the same order used in SRFI 43's vector-copy!`. Note that SRFI 43 will be part of the large language.
Vote yes to eliminate and rename as proposed, and no to leave as-is.
There were many complaints against the existing API, we should simplify this.
From Per Bothner:
Copying a slice from one vector/string into another is such a fundamental operation that it should be added, IMO, considering that it's tedious to write if "by hand", and that a standard library routine is likely to be much more efficient (especially for strings, since that avoids the need for boxing and unboxing the characters). [JC: Many implementations represent characters as immediates, however.]
One could also argue that "character" operations don't really make semantic sense in a Unicode world, and so string-set! has limited usefulness. Thus string-copy [with start/end arguments] and string-copy! are the actual useful "primitive" operations.
JC: These would be the five-argument versions based on the current bytevector-copy-partial!, possibly with renumbering of arguments depending on the outcome of #325.
Vote yes to add these destructive operations as proposed.
list-copy! is very rarely used and usually indicative of a broken algorithm. string-copy! is a step away from immutable strings, and terrible performance for utf8 implementations. vector-copy! is more useful but probably better left to a general vector library, though I could be convinced to include just it.
Currently there is no guarantee of this. Obviously the string->number only applies to the case where the radix is 10 or specified.
Specifying same is problematic in the presence of batch compilation
tower.
Basically all we can and should say is that string->number with a radix of 10 should behave the same as read.
R6RS changed these names to the more sensible exact and inexact. We need to decide if we want to follow suit, or provide both names, or write a disclaimer.
Vote r6rs for the short names, r5rs for the long names, or both for both.
We have a library system now, we shouldn't be too afraid to clean up names. The (scheme r5rs) library can be voted in if we want easy 100% compatibility.
The (scheme ieee) library exports the standard identifiers of IEEE 1178-1990. By my current reckoning, those identifiers are as follows:
`- * / + < <= = > >= abs acos and angle append apply asin assoc assq assv atan begin boolean? call-with-current-continuation car case cdr ceiling char->integer char-alphabetic? char-ci<? char-ci<=? char-ci=? char-ci>? char-ci>=? char-downcase char-lower-case? char-numeric? char-upcase char-upper-case? char-whitespace? char? char<? char<=? char=? char>? char>=? close-input-port close-output-port complex? cond cons cos current-input-port current-output-port define denominator display do eof-object? eq? equal? eqv? even? exact->inexact exact? exp expt floor for-each gcd if imag-part inexact->exact inexact? input-port? integer->char integer? lambda lcm length let let* letrec list list-ref list? log magnitude make-polar make-rectangular make-string make-vector map max member memq memv min modulo negative? newline not null? number->string number? numerator odd? open-input-file open-output-file or output-port? pair? peek-char positive? procedure? quasiquote quote quotient rational? rationalize read read-char real-part real? remainder reverse round set-car! set-cdr! set! sin sqrt string string->number string->symbol string-append string-ci<? string-ci<=? string-ci=? string-ci>? string-ci>=? string-length string-ref string-set! string? string<? string<=? string=? string>? string>=? substring symbol->string symbol? tan truncate vector vector-length vector-ref vector-set! vector? write write-char zero?`
As with any library other than (scheme base), implementations SHOULD (rather than MUST) provide this.
Vote yes to add this library.
I don't think people would ever want to use this instead of (scheme r5rs).
The (scheme r5rs) library exports the standard identifiers of R5RS Scheme other than transcript-{on,off}. By my current reckoning, those identifiers are as follows:
`- * / + < <= = > >= abs acos and angle append apply asin assoc assq assv atan begin boolean? call-with-current-continuation call-with-values car case cdr ceiling char->integer char-alphabetic? char-ci<? char-ci<=? char-ci=? char-ci>? char-ci>=? char-downcase char-lower-case? char-numeric? char-ready? char-upcase char-upper-case? char-whitespace? char? char<? char<=? char=? char>? char>=? close-input-port close-output-port complex? cond cons cos current-input-port current-output-port define define-syntax delay denominator display do dynamic-wind eof-object? eq? equal? eqv? eval even? exact->inexact exact? exp expt floor for-each force gcd if imag-part inexact->exact inexact? input-port? integer->char integer? interaction-environment lambda lcm length let let-syntax let* letrec letrec-syntax list list->string list->vector list-ref list-tail list? load log magnitude make-polar make-rectangular make-string make-vector map max member memq memv min modulo negative? newline not null-environment null? number->string number? numerator odd? open-input-file open-output-file or output-port? pair? peek-char positive? procedure? quasiquote quote quotient rational? rationalize read read-char real-part real? remainder reverse round scheme-report-environment set-car! set-cdr! set! sin sqrt string string->list string->number string->symbol string-append string-ci<? string-ci<=? string-ci=? string-ci>? string-ci>=? string-copy string-fill! string-length string-ref string-set! string? string<? string<=? string=? string>? string>=? substring symbol->string symbol? tan truncate values vector vector->list vector-fill! vector-length vector-ref vector-set! vector? with-input-from-file with-output-to-file write write-char zero?`
As with any library other than (scheme base), implementations SHOULD (rather than MUST) provide this. A disclaimer will be added that the semantics may not be exactly the same.
Vote yes to add this library.
This eases backwards-compatibility greatly.
The (scheme r6rs base) library exports the standard identifiers of the base library of R6RS. By my current reckoning, those identifiers are as follows:
`- * / + < <= = > >= abs acos and angle append apply asin atan begin boolean? call/cc call-with-current-continuation call-with-values car case cdr ceiling char? char<? char<=? char=? char>? char>=? char->integer complex? cond cons cos define define-syntax denominator dynamic-wind eq? equal? eqv? even? exact exact? exact-integer-sqrt exp expt finite? floor for-each gcd guard if imag-part import inexact inexact? integer? integer->char lambda lcm length let let* let*-values letrec letrec* letrec-syntax let-syntax let-values list list? list->string list->vector list-ref list-tail log magnitude make-polar make-rectangular make-string make-vector map max min nan? negative? not null? number? number->string numerator odd? or pair? positive? procedure? quasiquote quote rational? rationalize real? real-part reverse round set! sin sqrt string string? string<? string<=? string=? string>? string>=? string->list string->number string->symbol string-append string-copy string-for-each string-length string-ref substring symbol? symbol->string tan truncate values vector vector? vector->list vector-fill! vector-for-each vector-length vector-map vector-ref vector-set! zero?`
As with any library other than (scheme base), implementations SHOULD (rather than MUST) provide this. Full compliance will depend on voting for the procedures *-valued, assert, boolean=?, symbol=?. A disclaimer will be added that the semantics will not be exactly the same.
Vote yes to add this library.
I don't agree with including all of the R6RS base exports in the small language, and the small language has no obligation to support R6RS - that's the reason we have a large language.
Currently, to export my:foo and my:bar as foo and bar, one must write (export (rename my:foo foo) (rename my:bar bar)). This proposal allows (export (rename (my:foo foo) (my:bar bar))). This is incompatible with R6RS, but compatible with the rename sub-form of import.
Vote multiple to allow multiple renames in one rename clause as with the import version, r6rs to allow the R6RS-compatible syntax in the current draft, or both to allow both forms.
I like the simplicity of (length export-list) indicating the number of exports.
It's already a requirement that an eof-object cannot have an external representation, which means it cannot be any of the basic types in Section 3.2 except procedure or port. This is very improbable, and in fact none of my 40 test Schemes returns either a procedure or a port.
Doing this would allow eof-object? to be added to the list of disjoint type predicates in Section 3.2.
Vote yes to explicitly list the eof-object as a separate disjoint type.
This is a useful guarantee and a de-facto standard.
Specifically R7RS, IEEE-float, full-Unicode, Windows, POSIX, Unix, Darwin, Linux, BSD, FreeBSD, Solaris, PPC, SPARC, JVM, CLR, LLVM, ILP32, LP64, ILP64.
Note this is incompatible with existing implementations which provide these features. The correct case can often be ambiguous, and it's easiest to keep everything consistently lower case.
Vote mixed for mixed case and lower for lower case.
We have enough of a de-facto standard with existing implementation features, and as a general rule I hate mixed-case identifiers. People are inconsistent as to when they use it in different libraries (is it "utf8" or "UTF8"?, etc.), so it's easier to remember if everything is lowercase. It also saves having to touch the shift key.
If an exception is caught and leaves the current dynamic extent, obviously the after thunk must be run, but an uncaught exception has no semantics and is basically reverting to "is an error" semantics,
Possibly we should tighten this up in the standard, i.e. specify that there is a default exception handler which enters a continuation outside the extent of the whole program before exiting.
Vote unwind to specify that there is a default exception handler which leaves the current dynamic extent causing a full unwind (and thus forbidding a debugger), exit to specify that (modulo any diagnostic information) the program must simply exit without unwinding, or unspecified to leave this as is.
The default behavior in some existing implementations such as Gambit is to drop into a debugger, even in a batch program, so I don't think we can specify this.
Currently the report is silent about whether dynamic-wind handlers are invoked when exit is called.
The options are the same as in #335 above.
Contrary to #335 above, there is no reason not to unwind here. Once finalizers are supported (maybe in the large language) we'd probably want to require them to be run as well. Note this does mean that exit can't be a simple wrapper around the syscall.
eof-object returns an object which answers #t to eof-object?. This procedure is present in R6RS, where it must return the unique end-of-file object; that is not required here.
From Vincent Manis:
This isn't just an attempt to create a vain orthogonality; there are good reasons why arbitrary code might wish to return an eof object. For example, a DBMS interface might have a routine that returns one row, as a list or a vector, at a time; after the last, it is perfectly reasonable to return an eof object.
An argument against providing this is that the constructor may be trivially written, as shown [below]. A similar argument could be applied to zero?, newline, quotient, remainder, and modulo, among others. R7RS is not afraid to provide easy-to-implement procedures in the name of simplicity, orthogonality, or historical compatibility. The lack of an eof constructor is worth remedying.
(let* ((p (open-input-string "")) (x (read p))) (close-port p) x)Vote eof-object for a procedure of that name, or none to not add any such procedure.
I end up generating this often enough anyway with (read-char (open-input-string "")).
Currently the identifiers in library names can be any identifier. Under this proposal, the identifiers must not include any of `| \ ?* < " : > + [ ] /` or control characters after escapes are expanded.
If this proposal fails, its content will be included non-normatively as a should not.
Vote yes to restrict with must not.
R6RS had enough headaches with this and ":", we should avoid the same mistake.
Libraries do not necessarily have any mapping to files, nor does an implementation necessarily run on a system with a filesystem, however for those implementations which do so it may be worth adding such a note.
A library file contains a single library. A library named (A1 A2 AN) is in a file named "A1/A2/AN.sld" ("sld" for "Scheme Library Definition" or some other standardized file extension), relative to some "library path". For portability, library component names should be integers or lower-case identifiers that avoid certain prohibited characters. When a library or top-level imports some other library, the corresponding file is found in the obvious way.
Alternately, this can be left entirely to WG2 and/or packaging systems such as Snow.
Vote yes to add such a note or no to leave it out.
This belongs in Snow.
It is currently an error to attempt to import the same identifier from more than one library into another library or a top-level program, even if the identifier is not used anywhere in the new library or program. That requires programmers to make an arbitrary decision to exclude it from one library or the other.
Vote yes to agree with this proposal to require that, within a single static library (not with the environment procedure where any identifier may be subsequently used), an implementation must allow such multiple imports if the identifier is not referenced and does not occur in a syntax-rules template (which introduces conflicts with low-level macros introduced by WG2).
I think this is too difficult to specify and implement.
Currently, read-bytevector and read-bytevector! return an EOF object at EOF; otherwise, read-bytevector returns a non-empty bytevector and read-bytevector! returns the number of bytes read. Returning #u8() and 0, respectively, at EOF instead would make the results always the same type. This change would introduce the ambiguity that one would not be able to detect EOF when reading a bytevector of length 0 (which is to say, not reading any bytes at all).
Vote zero to return #u8() and 0 as in the proposal, and eof-object to return the eof-object as in the current draft. Vote zero! to make the change only for read-bytevector!.
Returning the eof-object as in the current draft is ubambiguous and consistent with R6RS.
All Scheme standards up to and including R6RS and R7RS draft-6 have consistently placed the full domain at the beginning of each entry. In most cases the domain consists only of the implicit type restrictions from the prototype, but in some cases there are additional domain restrictions that cannot be conveniently included in the prototype such as the following map restrictions:
It is an error if proc does not accept as many arguments as there are lists and return a single value.
It has been suggested to move this to an appropriate later point in the entry, to put more emphasis on the initial entry description. This has the disadvantage of splitting the domain into two places, which can more easily cause oversights and make quick domain confirmations difficult.
An alternative is to separate the additional domain restrictions from the initial description, as a separate short paragraph immediately following the prototype and possibly de-emphasized by making it smaller. his would keep the domain in one place and still allow let the first line of the description stand out prominently in the initial paragraph.
Vote start for the status quo, start-split for the separate de-emphasized option, or later to move additional restrictions to a later point.
I want to keep all domain information in a single place - I often know what a procedure does but need to double check only the domain, and having to jump around to find it is inconvenient.
Currently, the draft report implies that 0.0 and -0.0 must be the same in the sense of eqv?, because eqv? defers to = for numbers (with the possible exception of NaNs).
Vote same for the status quo, different to change to "must be different", or unspecified to change to "may be different".
This should fall out naturally from the definition of eqv?, which currently makes it clear that they must be the same, though we'll be revisiting this.
Currently, R7RS (tracking R5RS) does not constrain the sizes of exact integers beyond being required to represent the indices of strings, vectors and bytevectors.
R6RS requires systems to support "practically unlimited" size exact integers. It also requires that a subset of these exist, called fixnums, which must support at least the range -223 to 223-1. (All practical Schemes have larger ranges for their fixnums). This proposal suggests that we adopt this range as the minimum range of R7RS exact integers.
The immediate issue here is that a library name may contain (non-negative) exact integers as well as identifiers in R7RS. For such names to be portable, there must be a portable range of exact integers.
See FixnumInfo to see what 39 existing Schemes do.
Vote 24 to require 24 bits of precision, or none to leave this entirely unspecified.
R7RS small does not make the ugly "fixnum" distinction, and I don't see why we should set such arbitrary minimums, or forbid simplistic Scheme implementations on 16-bit machines.
We define mutating imports to be an error, however the standard currently says nothing about what happens when an exported binding is mutated from within the library where it's defined. In many common library implementations there will be no effect (i.e. the import effectively gets a copy of the original), whereas in a namespace based implementation the change will be reflected, so a conservative approach is to add a note saying the result is unspecified.
Vote shared to force the binding to be shared and the change reflected everywhere it's imported, separate to force the binding to be separate, none to make no comment, and unspecified or error to add a clarification to the standard to that effect.
This was the original intention, and leaving it out was an oversight.
A formal comment has proposed changing the epoch of current-second to 1970-01-01 00:00:00 TAI rather than 1970-01-01 00:00:10 TAI (00:00:00 UTC).
The actual time systems are independent of an epoch - the epoch is just convenient for computer systems.
The UTC-centric epoch was chosen (despite the use of TAI time) mostly because it is used in popular TAI times such as libtai and Olson's time library.
See http://lists.scheme-reports.org/pipermail/scheme-reports/2012-March/001943.html for more details.
Vote utc for the current draft's start-of-1970-in-utc epoch, or tai for the proposed start-of-1970-in-tai epoch.
I actually haven't been able to find any libraries which use the TAI epoch.