Skip to content
Find file
Fetching contributors…
Cannot retrieve contributors at this time
1301 lines (1157 sloc) 53.3 KB
Bugs can be reported on the help mailing list
or on the development mailing list
Please include enough information in a bug report that someone reading
it can reproduce the problem, i.e. don't write
Subject: apparent bug in PRINT-OBJECT (or *PRINT-LENGTH*?)
PRINT-OBJECT doesn't seem to work with *PRINT-LENGTH*. Is this a bug?
but instead
Subject: apparent bug in PRINT-OBJECT (or *PRINT-LENGTH*?)
In sbcl-1.2.3 running under OpenBSD 4.5 on my Alpha box, when
I compile and load the file
(PRINT X Y)))))
X Y)
then at the command line type
the program loops endlessly instead of printing the object.
There is also some information on bugs in the manual page and
in the TODO file. Eventually more such information may move here.
The gaps in the number sequence belong to old bug descriptions which
have gone away (typically because they were fixed, but sometimes for
other reasons, e.g. because they were moved elsewhere).
DEFSTRUCT almost certainly should overwrite the old LAYOUT information
instead of just punting when a contradictory structure definition
is loaded. As it is, if you redefine DEFSTRUCTs in a way which
changes their layout, you probably have to rebuild your entire
program, even if you know or guess enough about the internals of
SBCL to wager that this (undefined in ANSI) operation would be safe.
3: "type checking of structure slots"
ANSI specifies that a type mismatch in a structure slot
initialization value should not cause a warning.
This one might not be fixed for a while because while we're big
believers in ANSI compatibility and all, (1) there's no obvious
simple way to do it (short of disabling all warnings for type
mismatches everywhere), and (2) there's a good portable
workaround, and (3) by their own reasoning, it looks as though
ANSI may have gotten it wrong. ANSI justifies this specification
by saying
The restriction against issuing a warning for type mismatches
between a slot-initform and the corresponding slot's :TYPE
option is necessary because a slot-initform must be specified
in order to specify slot options; in some cases, no suitable
default may exist.
However, in SBCL (as in CMU CL or, for that matter, any compiler
which really understands Common Lisp types) a suitable default
does exist, in all cases, because the compiler understands the
concept of functions which never return (i.e. has return type NIL).
Thus, as a portable workaround, you can use a call to some
known-never-to-return function as the default. E.g.
(BAR (ERROR "missing :BAR argument")
(DEFUN REQUIRED-ARG () ; workaround for SBCL non-ANSI slot init typing
(ERROR "missing required argument"))
Such code should compile without complaint and work correctly either
on SBCL or on any other completely compliant Common Lisp system.
b: &AUX argument in a boa-constructor without a default value means
"do not initilize this slot" and does not cause type error. But
an error may be signalled at read time and it would be good if
SBCL did it.
c: Reading of not initialized slot sometimes causes SEGV.
(declaim (optimize (safety 3) (speed 1) (space 1)))
(defstruct foo
x y)
(defstruct (stringwise-foo (:include foo
(x "x" :type simple-string)
(y "y" :type simple-string))))
(defparameter *stringwise-foo*
(setf (foo-x *stringwise-foo*) 0)
(defun frob-stringwise-foo (sf)
(aref (stringwise-foo-x sf) 0))
(frob-stringwise-foo *stringwise-foo*)
The "compiling top-level form:" output ought to be condensed.
Perhaps any number of such consecutive lines ought to turn into a
single "compiling top-level forms:" line.
The way that the compiler munges types with arguments together
with types with no arguments (in e.g. TYPE-EXPAND) leads to
weirdness visible to the user:
(TYPEP 11 'FOO) => T
(TYPEP 11 '(FOO)) => T, which seems weird
(TYPEP 11 '(FIXNUM)) signals an error, as it should
The situation is complicated by the presence of Common Lisp types
like UNSIGNED-BYTE (which can either be used in list form or alone)
so I'm not 100% sure that the behavior above is actually illegal.
But I'm 90+% sure, and the following related behavior,
(TYPEP 11 'AND) => T
treating the bare symbol AND as equivalent to '(AND), is specifically
forbidden (by the ANSI specification of the AND type).
It would be nice if the
caught ERROR:
(during macroexpansion)
said what macroexpansion was at fault, e.g.
caught ERROR:
(during macroexpansion of IN-PACKAGE,
during macroexpansion of DEFFOO)
(Also, when this is fixed, we can enable the code in PROCLAIM which
checks for incompatible FTYPE redeclarations.)
(I *think* this is a bug. It certainly seems like strange behavior. But
the ANSI spec is scary, dark, and deep.. -- WHN)
(FORMAT NIL "~,1G" 1.4) => "1. "
(FORMAT NIL "~3,1G" 1.4) => "1. "
Sometimes (SB-EXT:QUIT) fails with
Argh! maximum interrupt nesting depth (4096) exceeded, exiting
Process inferior-lisp exited abnormally with code 1
I haven't noticed a repeatable case of this yet.
The printer doesn't report closures very well. This is true in
CMU CL 18b as well:
#<Closure Over Function "DEFUN STRUCTURE-SLOT-ACCESSOR" {134D1A1}>
It would be nice to make closures have a settable name slot,
and make things like DEFSTRUCT and FLET, which create closures,
set helpful values into this slot.
And as long as we're wishing, it would be awfully nice if INSPECT could
also report on closures, telling about the values of the bound variables.
The compiler assumes that any time a function of declared FTYPE
doesn't signal an error, its arguments were of the declared type.
E.g. compiling and loading
(COND ((> (FACTORIAL X) 1.0E6)
(FORMAT T "too big~%"))
(FORMAT T "exactly ~S~%" (FACTORIAL X)))
(FORMAT T "approximately ~S~%" (FACTORIAL X)))))
then executing
(FOO 1.5)
will cause the INTEGERP case to be selected, giving bogus output a la
exactly 2.5
This violates the "declarations are assertions" principle.
According to the ANSI spec, in the section "System Class FUNCTION",
this is a case of "lying to the compiler", but the lying is done
by the code which calls FACTORIAL with non-UNSIGNED-BYTE arguments,
not by the unexpectedly general definition of FACTORIAL. In any case,
"declarations are assertions" means that lying to the compiler should
cause an error to be signalled, and should not cause a bogus
result to be returned. Thus, the compiler should not assume
that arbitrary functions check their argument types. (It might
make sense to add another flag (CHECKED?) to DEFKNOWN to
identify functions which *do* check their argument types.)
(Also, verify that the compiler handles declared function
return types as assertions.)
TYPEP of VALUES types is sometimes implemented very inefficiently, e.g. in
where the implementation of the type check in function FOO
includes a full call to %TYPEP. There are also some fundamental problems
with the interpretation of VALUES types (inherited from CMU CL, and
from the ANSI CL standard) as discussed on the
mailing list, e.g. in Robert Maclachlan's post of 21 Jun 2000.
The definitions of SIGCONTEXT-FLOAT-REGISTER and
%SET-SIGCONTEXT-FLOAT-REGISTER in x86-vm.lisp say they're not
supported on FreeBSD because the floating point state is not saved,
but at least as of FreeBSD 4.0, the floating point state *is* saved,
so they could be supported after all. Very likely
SIGCONTEXT-FLOATING-POINT-MODES could now be supported, too.
(as discussed by Douglas Crosher on the cmucl-imp mailing list ca.
Aug. 10, 2000): CMUCL currently interprets 'member as '(member); same
issue with 'union, 'and, 'or etc. So even though according to the
ANSI spec, bare 'MEMBER, 'AND, and 'OR are not legal types, CMUCL
(and now SBCL) interpret them as legal types.
a slew of floating-point-related errors reported by Peter Van Eynde
on July 25, 2000:
b: SBCL's value for LEAST-POSITIVE-SHORT-FLOAT on the x86 is
bogus, and should probably be 1.4012985e-45. In SBCL,
(/ LEAST-POSITIVE-SHORT-FLOAT 2) returns a number smaller
than LEAST-POSITIVE-SHORT-FLOAT. Similar problems
c: Many expressions generate floating infinity on x86/Linux:
(/ 1 0.0)
(/ 1 0.0d0)
(EXPT 10.0 1000)
(EXPT 10.0d0 1000)
PVE's regression tests want them to raise errors. sbcl-
on x86/Linux generates the infinities instead. That might or
might not be conforming behavior, but it's also inconsistent,
which is almost certainly wrong. (Inconsistency: (/ 1 0.0)
should give the same result as (/ 1.0 0.0), but instead (/ 1 0.0)
generates SINGLE-FLOAT-POSITIVE-INFINITY and (/ 1.0 0.0)
signals an error.
d: (in section12.erg) various forms a la
don't give the right behavior.
type safety errors reported by Peter Van Eynde July 25, 2000:
c: (COERCE 'AND 'FUNCTION) returns something related to
(MACRO-FUNCTION 'AND), but ANSI says it should raise an error.
k: READ-BYTE is supposed to signal TYPE-ERROR when its argument is
not a binary input stream, but instead cheerfully reads from
character streams, e.g. (MAKE-STRING-INPUT-STREAM "abc").
The debugger LIST-LOCATIONS command doesn't work properly.
Compiling and loading
(FAIL 12)
then requesting a BACKTRACE at the debugger prompt gives no information
about where in the user program the problem occurred.
Paul Werkowski wrote on 2000-11-15
I am looking into this problem that showed up on the cmucl-help
list. It seems to me that the "implementation specific environment
hacking functions" found in pcl/walker.lisp are completely messed
up. The good thing is that they appear to be barely used within
PCL and the munged environment object is passed to cmucl only
in calls to macroexpand-1, which is probably why this case fails.
SBCL uses essentially the same code, so if the environment hacking
is screwed up, it affects us too.
Using the pretty-printer from the command prompt gives funny
results, apparently because the pretty-printer doesn't know
about user's command input, including the user's carriage return
that the user, and therefore the pretty-printer thinks that
the new output block should start indented 2 or more characters
rightward of the correct location.
As reported by Winton Davies on a CMU CL mailing list 2000-01-10,
and reported for SBCL by Martin Atzmueller 2000-10-20: (TRACE GETHASH)
crashes SBCL. In general tracing anything which is used in the
implementation of TRACE is likely to have the same problem.
As reported by Martin Atzmueller on sbcl-devel 26 Dec 2000,
ANSI says that WITH-OUTPUT-TO-STRING should have a keyword
:ELEMENT-TYPE, but in sbcl-0.6.9 this is not defined for
ANSI says in one place that type declarations can be abbreviated even
when the type name is not a symbol, e.g.
SBCL doesn't support this. But ANSI says in another place that this
isn't allowed. So it's not clear this is a bug after all. (See the
e-mail on on 2001-01-16 and 2001-01-17 from WHN
and Pierre Mai.)
as pointed out by Dan Barlow on sbcl-devel 2000-07-02:
an easily guessable temporary filename in a way which might open
applications using LOAD-FOREIGN to hijacking by malicious users
on the same machine. Incantations for doing this safely are
floating around the net in various "how to write secure programs
despite Unix" documents, and it would be good to (1) fix this in
LOAD-FOREIGN, and (2) hunt for any other code which uses temporary
files and make it share the same new safe logic.
(partially alleviated in sbcl- by a fix by Matthew Danish to
make the temporary filename less easily guessable)
RANDOM-INTEGER-EXTRA-BITS=10 may not be large enough for the RANDOM
RNG to be high quality near RANDOM-FIXNUM-MAX; it looks as though
the mean of the distribution can be systematically O(0.1%) wrong.
Just increasing R-I-E-B is probably not a good solution, since
it would decrease efficiency more than is probably necessary. Perhaps
using some sort of accept/reject method would be better.
Internally the compiler sometimes evaluates
(sb-kernel:type/= (specifier-type '*) (specifier-type t))
(I stumbled across this when I added an
(assert (not (eq type1 *wild-type*)))
in the NAMED :SIMPLE-= type method.) '* isn't really a type, and
in a type context should probably be translated to T, and so it's
probably wrong to ask whether it's equal to the T type and then (using
the EQ type comparison in the NAMED :SIMPLE-= type method) return NIL.
(I haven't tried to investigate this bug enough to guess whether
there might be any user-level symptoms.)
In fact, the type system is likely to depend on this inequality not
holding... * is not equivalent to T in many cases, such as
Inconsistencies between derived and declared VALUES return types for
DEFUN aren't checked very well. E.g. the logic which successfully
catches problems like
(declaim (ftype (function (fixnum) float) foo))
(defun foo (x)
(declare (type integer x))
(values x)) ; wrong return type, detected, gives warning, good!
fails to catch
(declaim (ftype (function (t) (values t t)) bar))
(defun bar (x)
(values x)) ; wrong number of return values, no warning, bad!
The cause of this is seems to be that (1) the internal function
VALUES-TYPES-EQUAL-OR-INTERSECT used to make the check handles its
arguments symmetrically, and (2) when the type checking code was
written back when when SBCL's code was still CMU CL, the intent
was that this case
(declaim (ftype (function (t) t) bar))
(defun bar (x)
(values x x)) ; wrong number of return values; should give warning?
not be warned for, because a two-valued return value is considered
to be compatible with callers who expects a single value to be
returned. That intent is probably not appropriate for modern ANSI
Common Lisp, but fixing this might be complicated because of other
divergences between auld-style and new-style handling of
multiple-VALUES types. (Some issues related to this were discussed
on cmucl-imp at some length sometime in 2000.)
The facility for dumping a running Lisp image to disk gets confused
when run without the PURIFY option, and creates an unnecessarily large
core file (apparently representing memory usage up to the previous
high-water mark). Moreover, when the file is loaded, it confuses the
GC, so that thereafter memory usage can never be reduced below that
In sbcl- (and in all earlier SBCL, and in CMU
CL), out-of-line structure slot setters are horribly inefficient
whenever the type of the slot is declared, because out-of-line
structure slot setters are implemented as closures to save space,
so the compiler doesn't compile the type test into code, but
instead just saves the type in a lexical closure and interprets it
at runtime.
A proper solution involves deciding whether it's really worth
saving space by implementing structure slot accessors as closures.
(If it's not worth it, the problem vanishes automatically. If it
is worth it, there are hacks we could use to force type tests to
be compiled anyway, and even shared. E.g. we could implement
an EQUAL hash table mapping from types to compiled type tests,
and save the appropriate compiled type test as part of each lexical
closure; or we could make the lexical closures be placeholders
which overwrite their old definition as a lexical closure with
a new compiled definition the first time that they're called.)
As a workaround for the problem, #'(SETF FOO) expressions can
be replaced with (EFFICIENT-SETF-FUNCTION FOO), where
(defmacro efficient-setf-function (place-function-name)
(or #+sbcl (and (sb-impl::info :function :accessor-for place-function-name)
;; a workaround for the problem, encouraging the
;; inline expansion of the structure accessor, so
;; that the compiler can optimize its type test
(let ((new-value (gensym "NEW-VALUE-"))
(structure-value (gensym "STRUCTURE-VALUE-")))
`(lambda (,new-value ,structure-value)
(setf (,place-function-name ,structure-value)
;; no problem, can just use the ordinary expansion
`(function (setf ,place-function-name))))
There's apparently a bug in CEILING optimization which caused
Douglas Crosher to patch the CMU CL version. Martin Atzmueller
applied the patches to SBCL and they didn't seem to cause problems
(as reported sbcl-devel 2001-05-04). However, since the patches
modify nontrivial code which was apparently written incorrectly
the first time around, until regression tests are written I'm not
comfortable merging the patches in the CVS version of SBCL.
(TIME (ROOM T)) reports more than 200 Mbytes consed even for
a clean, just-started SBCL system. And it seems to be right:
(ROOM T) can bring a small computer to its knees for a *long*
time trying to GC afterwards. Surely there's some more economical
way to implement (ROOM T).
When the compiler inline expands functions, it may be that different
kinds of return values are generated from different code branches.
E.g. an inline expansion of POSITION generates integer results
from one branch, and NIL results from another. When that inline
expansion is used in a context where only one of those results
is acceptable, e.g.
(defun foo (x)
(aref *a1* (position x *a2*)))
and the compiler can't prove that the unacceptable branch is
never taken, then bogus type mismatch warnings can be generated.
If you need to suppress the type mismatch warnings, you can
suppress the inline expansion,
(defun foo (x)
#+sbcl (declare (notinline position)) ; to suppress bug 117 bogowarnings
(aref *a1* (position x *a2*)))
or, sometimes, suppress them by declaring the result to be of an
appropriate type,
(defun foo (x)
(aref *a1* (the integer (position x *a2*))))
This is not a new compiler problem in 0.7.0, but the new compiler
transforms for FIND, POSITION, FIND-IF, and POSITION-IF make it
more conspicuous. If you don't need performance from these functions,
and the bogus warnings are a nuisance for you, you can return to
your pre-0.7.0 state of grace with
#+sbcl (declaim (notinline find position find-if position-if)) ; bug 117..
as reported by Eric Marsden on 2001-08-14:
when of course it should be NIL. (He says it only fails for X86,
not SPARC; dunno about Alpha.)
Also, "the same problem exists for LONG-FLOAT-EPSILON,
for the -negative- the + is replaced by a - in the test)."
Raymond Toy comments that this is tricky on the X86 since its FPU
uses 80-bit precision internally.
Even in sbcl-0.pre7.x, which is supposed to be free of the old
non-ANSI behavior of treating the function return type inferred
from the current function definition as a declaration of the
return type from any function of that name, the return type of NIL
is attached to FOO in 120a above, and used to optimize code which
calls FOO.
As of version 0.pre7.14, SBCL's implementation of MACROLET makes
the entire lexical environment at the point of MACROLET available
in the bodies of the macroexpander functions. In particular, it
allows the function bodies (which run at compile time) to try to
access lexical variables (which are only defined at runtime).
It doesn't even issue a warning, which is bad.
The SBCL behavior arguably conforms to the ANSI spec (since the
spec says that the behavior is undefined, ergo anything conforms).
However, it would be better to issue a compile-time error.
Unfortunately I (WHN) don't see any simple way to detect this
condition in order to issue such an error, so for the meantime
SBCL just does this weird broken "conforming" thing.
The ANSI standard says, in the definition of the special operator
The macro-expansion functions defined by MACROLET are defined
in the lexical environment in which the MACROLET form appears.
Declarations and MACROLET and SYMBOL-MACROLET definitions affect
the local macro definitions in a MACROLET, but the consequences
are undefined if the local macro definitions reference any
local variable or function bindings that are visible in that
lexical environment.
Then it seems to contradict itself by giving the example
(defun foo (x flag)
(macrolet ((fudge (z)
;The parameters x and flag are not accessible
; at this point; a reference to flag would be to
; the global variable of that name.
` (if flag (* ,z ,z) ,z)))
;The parameters x and flag are accessible here.
(+ x
(fudge x)
(fudge (+ x 1)))))
The comment "a reference to flag would be to the global variable
of the same name" sounds like good behavior for the system to have.
but actual specification quoted above says that the actual behavior
is undefined.
(Since macroexpanders are defined in a restricted version
of the lexical environment, containing no lexical variables and
functions, which seems to conform to ANSI and CLtL2, but signalling
a STYLE-WARNING for references to variables similar to locals might
be a good thing.)
(as reported by Gabe Garza on cmucl-help 2001-09-21)
(defvar *tmp* 3)
(defun test-pred (x y)
(eq x y))
(defun test-case ()
(let* ((x *tmp*)
(func (lambda () x)))
(print (eq func func))
(print (test-pred func func))
(delete func (list func))))
Now calling (TEST-CASE) gives output
(#<FUNCTION {500A9EF9}>)
Evidently Python thinks of the lambda as a code transformation so
much that it forgets that it's also an object.
Ideally, uninterning a symbol would allow it, and its associated
FDEFINITION and PROCLAIM data, to be reclaimed by the GC. However,
at least as of sbcl-0.7.0, this isn't the case. Information about
FDEFINITIONs and PROCLAIMed properties is stored in globaldb.lisp
essentially in ordinary (non-weak) hash tables keyed by symbols.
Thus, once a system has an entry in this system, it tends to live
forever, even when it is uninterned and all other references to it
are lost.
141: "pretty printing and backquote"
* '``(FOO ,@',@S)
* (write '`(, .ala.) :readably t :pretty t)
(note the space between the comma and the point)
(reported by Jesse Bouwman 2001-10-24 through the unfortunately
prominent SourceForge web/db bug tracking system, which is
unfortunately not a reliable way to get a timely response from
the SBCL maintainers)
In the course of trying to build a test case for an
application error, I encountered this behavior:
If you start up sbcl, and then lay on CTRL-C for a
minute or two, the lisp process will eventually say:
%PRIMITIVE HALT called; the party is over.
and throw you into the monitor. If I start up lisp,
attach to the process with strace, and then do the same
(abusive) thing, I get instead:
access failure in heap page not marked as write-protected
and the monitor again. I don't know enough to have the
faintest idea of what is going on here.
This is with sbcl 6.12, uname -a reports:
Linux prep 2.2.19 #4 SMP Tue Apr 24 13:59:52 CDT 2001 i686 unknown
I (WHN) have verified that the same thing occurs on sbcl-0.pre7.141
under OpenBSD 2.9 on my X86 laptop. Do be patient when you try it:
it took more than two minutes (but less than five) for me.
(This was once known as IR1-4, but it lived on even after the
IR1 interpreter went to the big bit bucket in the sky.)
The system accepts DECLAIM in most places where DECLARE would be
accepted, without even issuing a warning. ANSI allows this, but since
it's fairly easy to mistype DECLAIM instead of DECLARE, and the
meaning is rather different, and it's unlikely that the user
has a good reason for doing DECLAIM not at top level, it would be
good to issue a STYLE-WARNING when this happens. A possible
fix would be to issue STYLE-WARNINGs for DECLAIMs not at top level,
or perhaps to issue STYLE-WARNINGs for any EVAL-WHEN not at top level.
[This is considered an IR1-interpreter-related bug because until
EVAL-WHEN is rewritten, which won't happen until after the IR1
interpreter is gone, the system's notion of what's a top-level form
and what's not will remain too confused to fix this problem.]
ANSI allows types `(COMPLEX ,FOO) to use very hairy values for
COMPLEX implementation didn't deal with this, and hasn't been
upgraded to do so. (This doesn't seem to be a high priority
conformance problem, since seems hard to construct useful code
where it matters.)
Floating point errors are reported poorly. E.g. on x86 OpenBSD
with sbcl-0.7.1,
* (expt 2.0 12777)
debugger invoked on condition of type SB-KERNEL:FLOATING-POINT-EXCEPTION:
An arithmetic error SB-KERNEL:FLOATING-POINT-EXCEPTION was signalled.
No traps are enabled? How can this be?
It should be possible to be much more specific (overflow, division
by zero, etc.) and of course the "How can this be?" should be fixable.
See also bugs #45.c and #183
In sbcl- on x86, COMPILE-FILE on the file
(in-package :cl-user)
(defvar *thing*)
(defvar *zoom*)
(defstruct foo bar bletch)
(defun %zeep ()
(labels ((kidify1 (kid)
(kid-frob (kid)
(if *thing*
(setf sweptm
(m+ (frobnicate kid)
(kidify1 kid))))
(declare (inline kid-frob))
(map nil
(the simple-vector (foo-bar perd)))))
fails with
debugger invoked on condition of type TYPE-ERROR:
The value NIL is not of type SB-C::NODE.
The location of this failure has moved around as various related
issues were cleaned up. As of sbcl-, it occurs in
(Python LET-converts KIDIFY1 into KID-FROB, then tries to inline
expand KID-FROB into %ZEEP. Having partially done it, it sees a call
of KIDIFY1, which already does not exist. So it gives up on
expansion, leaving garbage consisting of infinished blocks of the
partially converted function.)
(due to reordering of the compiler this example is compiled
successfully by 0.7.14, but the bug probably remains)
(reported by Robert E. Brown 2002-04-16)
When a function is called with too few arguments, causing the
debugger to be entered, the uninitialized slots in the bad call frame
seem to cause GCish problems, being interpreted as tagged data even
though they're not. In particular, executing ROOM in the
debugger at that point causes AVER failures:
* (machine-type)
* (lisp-implementation-version)
* (typep 10)
0] (room)
(Christophe Rhodes reports that this doesn't occur on the SPARC, which
isn't too surprising since there are many differences in stack
implementation and GC conservatism between the X86 and other ports.)
In sbcl-, compiling the (illegal) code
(in-package :cl-user)
(defmethod prove ((uustk uustk))
(zap ((frob () nil))
gives the (not terribly clear) error message
; caught ERROR:
; (during macroexpansion of (DEFMETHOD PROVE ...))
; can't get template for (FROB NIL NIL)
The problem seems to be that the code walker used by the DEFMETHOD
macro is unhappy with the illegal syntax in the method body, and
is giving an unclear error message.
The compiler sometimes tries to constant-fold expressions before
it checks to see whether they can be reached. This can lead to
bogus warnings about errors in the constant folding, e.g. in code
(WRITE-STRING (> X 0) "+" "0"))
compiled in a context where the compiler can prove that X is NIL,
and the compiler complains that (> X 0) causes a type error because
NIL isn't a valid argument to #'>. Until sbcl- or so this
caused a full WARNING, which made the bug really annoying because then
COMPILE and COMPILE-FILE returned FAILURE-P=T for perfectly legal
code. Since then the warning has been downgraded to STYLE-WARNING,
so it's still a bug but at least it's a little less annoying.
183: "IEEE floating point issues"
Even where floating point handling is being dealt with relatively
well (as of sbcl-0.7.5, on sparc/sunos and alpha; see bug #146), the
accrued-exceptions and current-exceptions part of the fp control
word don't seem to bear much relation to reality. E.g. on
* (/ 1.0 0.0)
debugger invoked on condition of type DIVISION-BY-ZERO:
arithmetic error DIVISION-BY-ZERO signalled
0] (sb-vm::get-floating-point-modes)
0] abort
* (sb-vm::get-floating-point-modes)
187: "type inference confusion around DEFTRANSFORM time"
(reported even more verbosely on sbcl-devel 2002-06-28 as "strange
After the file below is compiled and loaded in sbcl-0.7.5, executing
at the REPL returns an adjustable vector, which is wrong. Presumably
somehow the DERIVE-TYPE information for the output values of %WAD is
being mispropagated as a type constraint on the input values of %WAD,
and so causing the type test to be optimized away. It's unclear how
hand-expanding the DEFTRANSFORM would change this, but it suggests
the DEFTRANSFORM machinery (or at least the way DEFTRANSFORMs are
invoked at a particular phase) is involved.
(cl:in-package :sb-c)
(eval-when (:compile-toplevel)
;;; standin for %DATA-VECTOR-AND-INDEX
(defknown %dvai (array index)
(values t t)
(foldable flushable))
(deftransform %dvai ((array index)
(vector t)
:important t)
(let* ((atype (continuation-type array))
(eltype (array-type-specialized-element-type atype)))
(when (eq eltype *wild-type*)
"specialized array element type not known at compile-time"))
(when (not (array-type-complexp atype))
(give-up-ir1-transform "SIMPLE array!"))
`(if (array-header-p array)
(%wad array index nil)
(values array index))))
;;; standin for %WITH-ARRAY-DATA
(defknown %wad (array index (or index null))
(values (simple-array * (*)) index index index)
(foldable flushable))
;;; (Commenting out this optimizer causes the bug to go away.)
(defoptimizer (%wad derive-type) ((array start end))
(let ((atype (continuation-type array)))
(when (array-type-p atype)
`(values (simple-array ,(type-specifier
(array-type-specialized-element-type atype))
index index index)))))
(defun %wad (array start end)
(format t "~&in %WAD~%")
(%with-array-data array start end))
(cl:in-package :cl-user)
(defun tcx (v i)
(declare (type (vector t) v))
(declare (notinline sb-kernel::%with-array-data))
;; (Hand-expending DEFTRANSFORM %DVAI here also causes the bug to
;; go away.)
(sb-c::%dvai v i))
188: "compiler performance fiasco involving type inference and UNION-TYPE"
(In sbcl-, DEFTRANSFORM CONCATENATE was commented out until this
bug could be fixed properly, so you won't see the bug unless you restore
the DEFTRANSFORM by hand.) In sbcl- on a 700 MHz Pentium III,
(time (compile
'(lambda ()
(declare (optimize (safety 3)))
(declare (optimize (compilation-speed 2)))
(declare (optimize (speed 1) (debug 1) (space 1)))
(let ((fn "if-this-file-exists-the-universe-is-strange"))
(load fn :if-does-not-exist nil)
(load (concatenate 'string fn ".lisp") :if-does-not-exist nil)
(load (concatenate 'string fn ".fasl") :if-does-not-exist nil)
(load (concatenate 'string fn ".misc-garbage")
:if-does-not-exist nil)))))
134.552 seconds of real time
133.35156 seconds of user run time
0.03125 seconds of system run time
[Run times include 2.787 seconds GC run time.]
0 page faults and
246883368 bytes consed.
BACKTRACE from Ctrl-C in the compilation shows that the compiler is
thinking about type relationships involving types like
(OR (INTEGER 576 576)
(INTEGER 1192 1192)
(INTEGER 2536 2536)
(INTEGER 1816 1816)
(INTEGER 2752 2752)
(INTEGER 1600 1600)
(INTEGER 2640 2640)
(INTEGER 1808 1808)
(INTEGER 1296 1296)
In recent SBCL the following example also illustrates this bug:
(time (compile
'(lambda ()
(declare (optimize (safety 3)))
(declare (optimize (compilation-speed 2)))
(declare (optimize (speed 1) (debug 1) (space 1)))
(let ((start 4))
(declare (type (integer 0) start))
(print (incf start 22))
(print (incf start 26))
(print (incf start 28)))
(let ((start 6))
(declare (type (integer 0) start))
(print (incf start 22))
(print (incf start 26)))
(let ((start 10))
(declare (type (integer 0) start))
(print (incf start 22))
(print (incf start 26))))))
190: "PPC/Linux pipe? buffer? bug"
In sbcl-0.7.6, the test script sometimes hangs
on the PPC/Linux platform, waiting for a zombie env process. This
is a classic symptom of buffer filling and deadlock, but it seems
only sporadically reproducible.
191: "Miscellaneous PCL deficiencies"
(reported by Alexey Dejneka sbcl-devel 2002-08-04)
a. DEFCLASS does not inform the compiler about generated
functions. Compiling a file with
results in a STYLE-WARNING:
APD's fix for this was checked in to sbcl-, but Pierre
Mai points out that the declamation of functions is in fact
incorrect in some cases (most notably for structure
classes). This means that at present erroneous attempts to use
WITH-SLOTS and the like on classes with metaclass STRUCTURE-CLASS
won't get the corresponding STYLE-WARNING.
c. the examples in CLHS (regarding generic function lambda
lists and &KEY arguments) do not signal errors when they should.
192: "Python treats free type declarations as promises."
b. What seemed like the same fundamental problem as bug 192a, but
was not fixed by the same (APD "more strict type checking
sbcl-devel 2002-08-97) patch:
(DOTIMES (I ...) (DOTIMES (J ...) (DECLARE ...) ...)):
(declaim (optimize (speed 1) (safety 3)))
(defun trust-assertion (i)
(dotimes (j i)
(declare (type (mod 4) i)) ; when commented out, behavior changes!
(unless (< i 5)
(print j))))
(trust-assertion 6) ; prints nothing unless DECLARE is commented out
(see bug 203)
c. (defun foo (x y)
(locally (declare (type fixnum x y))
(+ x (* 2 y))))
(foo 1.1 2) => 5.1
194: "no error from (THE REAL '(1 2 3)) in some cases"
fixed parts:
a. In sbcl-,
(multiple-value-prog1 (progn (the real '(1 2 3))))
returns (1 2 3) instead of signalling an error. This was fixed by
APD's "more strict type checking patch", but although the fixed
code (in sbcl- works (signals TYPE-ERROR) interactively,
it's difficult to write a regression test for it, because
still returns (1 2 3).
still-broken parts:
returns (1 2 3). (As above, this shows up when writing regression
tests for fixed-ness of part a.)
c. Also in sbcl-, (IGNORE-ERRORS (THE REAL '(1 2 3))) => (1 2 3).
d. At the REPL,
(null (ignore-errors
(let ((arg1 1)
(arg2 (identity (the real #(1 2 3)))))
(if (< arg1 arg2) arg1 arg2))))
=> T
but putting the same expression inside (DEFUN FOO () ...),
(FOO) => NIL.
* Actually this entry is probably multiple bugs, as
Alexey Dejneka commented on sbcl-devel 2002-09-03:)
I don't think that placing these two bugs in one entry is
a good idea: they have different explanations. The second
(min 1 nil) is caused by flushing of unused code--IDENTITY
can do nothing with it. So it is really bug 122. The first
(min nil) is due to M-V-PROG1: substituting a continuation
for the result, it forgets about type assertion. The purpose
of IDENTITY is to save the restricted continuation from
inaccurate transformations.
* Alexey Dejneka pointed out that
work as they should.
201: "Incautious type inference from compound CONS types"
(reported by APD sbcl-devel 2002-09-17)
(FORMAT NIL "~S IS ~S, Y = ~S"
(FOO ' (1 . 2)) => "NIL IS INTEGER, Y = 1"
Compiler does not check THEs on unused values, e.g. in
(progn (the real (list 1)) t)
This situation may appear during optimizing away degenerate cases of
certain functions: see bug 192b.
205: "environment issues in cross compiler"
(These bugs have no impact on user code, but should be fixed or
a. Macroexpanders introduced with MACROLET are defined in the null
lexical environment.
b. The body of (EVAL-WHEN (:COMPILE-TOPLEVEL) ...) is evaluated in
the null lexical environment.
c. The cross-compiler cannot inline functions defined in a non-null
lexical environment.
206: ":SB-FLUID feature broken"
(reported by Antonio Martinez-Shotton sbcl-devel 2002-10-07)
Enabling :SB-FLUID in the target-features list in sbcl-0.7.8 breaks
the build.
207: "poorly distributed SXHASH results for compound data"
SBCL's SXHASH could probably try a little harder. ANSI: "the
intent is that an implementation should make a good-faith
effort to produce hash-codes that are well distributed
within the range of non-negative fixnums". But
(let ((hits (make-hash-table)))
(dotimes (i 16)
(dotimes (j 16)
(let* ((ij (cons i j))
(newlist (push ij (gethash (sxhash ij) hits))))
(when (cdr newlist)
(format t "~&collision: ~S~%" newlist))))))
reports lots of collisions in sbcl-0.7.8. A stronger MIX function
would be an obvious way of fix. Maybe it would be acceptably efficient
to redo MIX using a lookup into a 256-entry s-box containing
29-bit pseudorandom numbers?
208: "package confusion in PCL handling of structure slot handlers"
In sbcl-0.7.8 compiling and loading
(in-package :cl)
(defstruct foo (slot (error "missing")) :type list :read-only t)
(defmethod print-object ((foo foo) stream) (print nil stream))
causes CERROR "attempting to modify a symbol in the COMMON-LISP
package: FOO-SLOT". (This is fairly bad code, but still it's hard
to see that it should cause symbols to be interned in the CL package.)
211: "keywords processing"
a. :ALLOW-OTHER-KEYS T should allow a function to receive an odd
number of keyword arguments.
e. Compiling
(flet ((foo (&key y) (list y)))
(list (foo :y 1 :y 2)))
issues confusing message
; (FOO :Y 1 :Y 2)
; The variable #:G15 is defined but never used.
212: "Sequence functions and circular arguments"
COERCE, MERGE and CONCATENATE go into an infinite loop when given
circular arguments; it would be good for the user if they could be
given an error instead (ANSI 17.1.1 allows this behaviour on the part
of the implementation, as conforming code cannot give non-proper
sequences to these functions. MAP also has this problem (and
solution), though arguably the convenience of being able to do
(MAP 'LIST '+ FOO '#1=(1 . #1#))
might be classed as more important (though signalling an error when
all of the arguments are circular is probably desireable).
213: "Sequence functions and type checking"
various complicated, though recognizeable, CONS types [e.g.
which according to ANSI should be recognized] (and, in SAFETY 3
code, should return a list of LENGTH 2 or signal an error)
b. MAP, when given a type argument that is SUBTYPEP LIST, does not
check that it will return a sequence of the given type. Fixing
it along the same lines as the others (cf. work done around
sbcl- is possible, but doing so efficiently didn't look
entirely straightforward.
c. All of these functions will silently accept a type of the form
whether or not the return value is of this type. This is
probably permitted by ANSI (see "Exceptional Situations" under
ANSI MAKE-SEQUENCE), but the DERIVE-TYPE mechanism does not
know about this escape clause, so code of the form
can erroneously return T.
SBCL fails to compile
(declare (optimize (inhibit-warnings 0) (compilation-speed 2)))
(flet ((foo (&key (x :vx x-p)) (list x x-p)))
(foo 1 2)))
or a more simple example:
(declare (optimize (inhibit-warnings 0) (compilation-speed 2)))
(lambda (x) (declare (fixnum x)) (if (< x 0) 0 (1- x))))
215: ":TEST-NOT handling by functions"
a. FIND and POSITION currently signal errors when given non-NIL for
both their :TEST and (deprecated) :TEST-NOT arguments, but by
ANSI 17.2 "the consequences are unspecified", which by ANSI 1.4.2
means that the effect is "unpredictable but harmless". It's not
clear what that actually means; it may preclude conforming
implementations from signalling errors.
b. COUNT, REMOVE and the like give priority to a :TEST-NOT argument
when conflict occurs. As a quality of implementation issue, it
might be preferable to treat :TEST and :TEST-NOT as being in some
sense the same &KEY, and effectively take the first test function in
the argument list.
c. Again, a quality of implementation issue: it would be good to issue a
STYLE-WARNING at compile-time for calls with :TEST-NOT, and a
WARNING for calls with both :TEST and :TEST-NOT; possibly this
latter should be WARNed about at execute-time too.
216: "debugger confused by frames with invalid number of arguments"
In sbcl-, executing e.g. (VECTOR-PUSH-EXTEND T), BACKTRACE, Q
leaves the system confused, enough so that (QUIT) no longer works.
It's as though the process of working with the uninitialized slot in
the bad VECTOR-PUSH-EXTEND frame causes GC problems, though that may
not be the actual problem. (CMU CL 18c doesn't have problems with this.)
217: "Bad type operations with FUNCTION types"
In sbcl.0.7.7:
* (values-type-union (specifier-type '(function (base-char)))
(specifier-type '(function (integer))))
It causes insertion of wrong type assertions into generated
code. E.g.
(defun foo (x s)
(let ((f (etypecase x
(character #'write-char)
(integer #'write-byte))))
(funcall f x s)
(etypecase x
(character (write-char x s))
(integer (write-byte x s)))))
Then (FOO #\1 *STANDARD-OUTPUT*) signals type error.
(In the result type is (FUNCTION * *), so Python does not
produce invalid code, but type checking is not accurate. Similar
problems exist with VALUES-TYPE-INTERSECTION.)
218: "VALUES type specifier semantics"
(THE (VALUES ...) ...) in safe code discards extra values.
(defun test (x y) (the (values integer) (truncate x y)))
(test 10 4) => 2
Sbcl 0.7.9 fails to compile
(multiple-value-call #'list
(the integer (helper))
Type check for INTEGER, the result of which serves as the first
argument of M-V-C, is inserted after evaluation of NIL. So arguments
of M-V-C are pushed in the wrong order. As a temporary workaround
type checking was disabled for M-V-Cs in A better solution
would be to put the check between evaluation of arguments, but it
could be tricky to check result types of PROG1, IF etc.
(subtypep 'function '(function)) => nil, t.
233: bugs in constraint propagation
(defun foo (x)
(declare (optimize (speed 2) (safety 3)))
(let ((y 0d0))
(the double-float x)
(setq y (+ x 1d0))
(setq x 3d0)
(quux y (+ y 2d0) (* y 3d0)))))
(foo 4) => segmentation violation
(see also bug 236)
(declaim (optimize (speed 2) (safety 3)))
(defun foo (x y)
(if (typep (prog1 x (setq x y)) 'double-float)
(+ x 1d0)
(+ x 2)))
(foo 1d0 5) => segmentation violation
235: "type system and inline expansion"
(declaim (ftype (function (cons) number) acc))
(declaim (inline acc))
(defun acc (c)
(the number (car c)))
(defun foo (x y)
(values (locally (declare (optimize (safety 0)))
(acc x))
(locally (declare (optimize (safety 3)))
(acc y))))
(foo '(nil) '(t)) => NIL, T.
b. (reported by brown on #lisp 2003-01-21)
(defun find-it (x)
(declare (optimize (speed 3) (safety 0)))
(declare (notinline mapcar))
(let ((z (mapcar #'car x)))
(find 'foobar z)))
Without (DECLARE (NOTINLINE MAPCAR)), Python cannot derive that Z is
236: "THE semantics is broken"
(defun foo (a f)
(declare (optimize (speed 2) (safety 0)))
(+ 1d0
(the double-float
(svref a 0)
(unless f (return-from foo 0))))))
(foo #(4) nil) => SEGV
VOP selection thinks that in unsafe code result type assertions
should be valid immediately. (See also bug 233a.)
The similar problem exists for TRULY-THE.
237: "Environment arguments to type functions"
UPGRADED-COMPLEX-PART-TYPE now have an optional environment
argument, but they ignore it completely. This is almost
certainly not correct.
b. Also, the compiler's optimizers for TYPEP have not been informed
about the new argument; consequently, they will not transform
calls of the form (TYPEP 1 'INTEGER NIL), even though this is
just as optimizeable as (TYPEP 1 'INTEGER).
238: "REPL compiler overenthusiasm for CLOS code"
From the REPL,
* (defclass foo () ())
* (defmethod bar ((x foo) (foo foo)) (call-next-method))
causes approximately 100 lines of code deletion notes. Some
discussion on this issue happened under the title 'Three "interesting"
bugs in PCL', resulting in a fix for this oververbosity from the
compiler proper; however, the problem persists in the interactor
because the notion of original source is not preserved: for the
compiler, the original source of the above expression is (DEFMETHOD
BAR ((X FOO) (FOO FOO)) (CALL-NEXT-METHOD)), while by the time the
compiler gets its hands on the code needing compilation from the REPL,
it has been macroexpanded several times.
A symptom of the same underlying problem, reported by Tony Martinez:
* (handler-case
(with-input-from-string (*query-io* " no")
(simple-type-error () 'error))
; note: deleting unreachable code
; compilation unit finished
; printed 1 note
241: "DEFCLASS mysteriously remembers uninterned accessor names."
(from tonyms on #lisp IRC 2003-02-25)
In sbcl-, typing
(defclass foo () ((bar :accessor foo-bar)))
(profile foo-bar)
(unintern 'foo-bar)
(defclass foo () ((bar :accessor foo-bar)))
gives the error message
"#:FOO-BAR already names an ordinary function or a macro."
So it's somehow checking the uninterned old accessor name instead
of the new requested accessor name, which seems broken to me (WHN).
242: "WRITE-SEQUENCE suboptimality"
(observed from clx performance)
In sbcl-0.7.13, WRITE-SEQUENCE of a sequence of type
(SIMPLE-ARRAY (UNSIGNED-BYTE 8) (*)) on a stream with element-type
(UNSIGNED-BYTE 8) will write to the stream one byte at a time,
rather than writing the sequence in one go, leading to severe
performance degradation.
243: "STYLE-WARNING overenthusiasm for unused variables"
(observed from clx compilation)
In sbcl-0.7.14, in the presence of the macros
somewhat surprising style warnings are emitted for
; in: LAMBDA (Y)
; (LAMBDA (Y) (FOO Y))
; The variable Y is defined but never used.
244: "optimizing away tests for &KEY args of type declared in DEFKNOWN"
(caught by clocc-ansi-test :EXCEPSIT-LEGACY-1050)
succeeds with no error (ignoring the bogus :EXTERNAL-FORMAT argument)
apparently because the test is optimized away. The problem doesn't
exist in sbcl-0.pre8.19. Deleting the (MEMBER :DEFAULT) declaration
for :EXTERNAL-FORMAT in DEFKNOWN OPEN (and LOAD) is a workaround for
the problem (and should be removed when the problem is fixed).
245: bugs in disassembler
a. On X86 an immediate operand for IMUL is printed incorrectly.
b. On X86 operand size prefix is not recognized.
246: "NTH-VALUE scaling problem"
NTH-VALUE's current implementation for constant integers scales in
compile-time as O(n^4), as indeed must the optional dispatch
mechanism on which it is implemented. While it is unlikely to
matter in real user code, it's still unpleasant to observe that
(NTH-VALUE 1000 (VALUES-LIST (MAKE-LIST 1001))) takes several hours
to compile.
248: "reporting errors in type specifier syntax"
(TYPEP 1 '(SYMBOL NIL)) says something about "unknown type
These labels were used for bugs related to the old IR1 interpreter.
The # values reached 6 before the category was closed down.
Jump to Line
Something went wrong with that request. Please try again.