+192: "Python treats free type declarations as promises."
+ a. original report by Alexey Dejneka (on sbcl-devel 2002-08-26):
+ (declaim (optimize (speed 0) (safety 3)))
+ (defun f (x)
+ (declare (real x))
+ (+ x
+ (locally (declare (single-float x))
+ (sin x))))
+ Now (F NIL) correctly gives a type error, but (F 100) gives
+ a segmentation violation.
+ b. same fundamental problem in a different way, easy to stumble
+ across if you mistype and declare the wrong index in
+ (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
+
+193: "unhelpful CLOS error reporting when the primary method is missing"
+ In sbcl-0.7.7, when
+ (defmethod foo :before ((x t)) (print x))
+ is the only method defined on FOO, the error reporting when e.g.
+ (foo 12)
+ is relatively unhelpful:
+ There is no primary method for the generic function
+ #<STANDARD-GENERIC-FUNCTION FOO (1)>.
+ with the offending argument nowhere visible in the backtrace. This
+ continues even if there *are* primary methods, just not for the
+ specified arg type, e.g.
+ (defmethod foo ((x character)) (print x))
+ (defmethod foo ((x string)) (print x))
+ (defmethod foo ((x pathname)) ...)
+ In that case it could be very helpful to know what argument value is
+ falling through the cracks of the defined primary methods, but the
+ error message stays the same (even BACKTRACE doesn't tell you what the
+ bad argument value is).
+
+194: "no error from (THE REAL '(1 2 3)) in some cases"
+ In sbcl-0.7.7.9,
+ (multiple-value-prog1 (progn (the real '(1 2 3))))
+ returns (1 2 3) instead of signalling an error. Also in sbcl-0.7.7.9,
+ a more complicated instance of this bug kept
+ (IGNORE-ERRORS (MIN '(1 2 3))) from returning NIL as it should when
+ the MIN source transform expanded to (THE REAL '(1 2 3)), because
+ (IGNORE-ERRORS (THE REAL '(1 2 3))) returns (1 2 3).
+ Alexey Dejneka pointed out that
+ (IGNORE-ERRORS (IDENTITY (THE REAL '(1 2 3)))) works as it should.
+ (IGNORE-ERRORS (VALUES (THE REAL '(1 2 3)))) also works as it should.
+ Perhaps this is another case of VALUES type intersections behaving
+ in non-useful ways?
+ When I (WHN) tried to use the VALUES trick to work around this bug
+ in the MIN source transform, it didn't work for
+ (assert (null (ignore-errors (min 1 #(1 2 3)))))
+ Hand-expanding the source transform, I get
+ (assert (null (ignore-errors
+ (let ((arg1 1)
+ (arg2 (identity (the real #(1 2 3)))))
+ (if (< arg1 arg2) arg1 arg2)))))
+ which fails (i.e. the assertion fails, because the IGNORE-ERRORS
+ doesn't report MIN signalling a type error). At the REPL
+ (null (ignore-errors
+ (let ((arg1 1)
+ (arg2 (identity (the real #(1 2 3)))))
+ (if (< arg1 arg2) arg1 arg2))))
+ => T
+ but when this expression is used as the body of (DEFUN FOO () ...)
+ then (FOO)=>NIL.
+
+195: "confusing reporting of not-a-REAL TYPE-ERRORs from THE REAL"
+ In sbcl-0.7.7.10, (THE REAL #(1 2 3)) signals a type error which
+ prints as "This is not a (OR SINGLE-FLOAT DOUBLE-FLOAT RATIONAL)".
+ The (OR SINGLE-FLOAT DOUBLE-FLOAT RATIONAL) representation of
+ REAL is unnecessarily confusing, especially since it relies on
+ internal implementation knowledge that even with SHORT-FLOAT
+ and LONG-FLOAT left out of the union, this type is equal to REAL.
+ So it'd be better just to say "This is not a REAL".
+
projects / sbcl.git / blobdiff