Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

ANSI-TEST fixes #261

Open
wants to merge 9 commits into
base: master
Choose a base branch
from
Open

ANSI-TEST fixes #261

wants to merge 9 commits into from

Conversation

phoe
Copy link
Contributor

@phoe phoe commented Dec 2, 2019

This is a series of commits that fixes multiple ANSI-TEST failures from the upstream ANSI-TEST.

Please review and test as appropriate. I have tested individual commits on Travis and also built CCL and ran ANSI-TEST on my local machine after cherry-picking the commits for this branch.

No bootstrapping issues are present in this MR.

The failing test's description:

;;; Test that SHIFTF returns a single value, even though the first
;;; place has multiple values.

The form:

(let ((x 'a) (y 'b))
  (values
   (multiple-value-list (shiftf (values x y) (floor 10 3)))
   x y))

Was expected to return the following three values:

(A) 3 1

But instead evaluated to:

(A B) 3 1
The failing test's description:

;;; &whole is followed by a destructuring pattern (see 3.4.4.1.2)

The form:

(macrolet ((%m (&whole (m a b) c d) `(quote (,m ,a ,b ,c ,d))))
  (%m 1 2)

Was expected to return the following value:

(%M 1 2 1 2)

But instead signaled an error, as MACROLET did not expect &WHOLE
to be followed by a destructuring pattern.
The failing test's description:

;;; Uninterned symbols must not contain a package prefix (see CLHS 2.4.8.5)

The form:

(read-from-string "#:a:b")

Was expected to signal a READER-ERROR, but signaled a SIMPLE-ERROR instead.
The body of the failing test:

(deftest compile-file.16
  (let* ((file #p"compile-file-test-file-5.lsp")
         (target-pathname (compile-file-pathname file))
         (*compile-print* nil)
         (*compile-verbose* nil))
    (when (probe-file target-pathname)
      (delete-file target-pathname))
    (compile-file file)
    (load target-pathname)
    (values
     (equalpt-or-report (truename file) (funcall 'compile-file-test-fun.5))
     (equalpt-or-report (pathname (merge-pathnames file))
                        (funcall 'compile-file-test-fun.5a))))
  t t)

Contents of file `compile-file-test-file-5.lsp`:

(in-package "CL-TEST")

(defun compile-file-test-fun.5 ()
  '#.*compile-file-truename*)

(defun compile-file-test-fun.5a ()
  '#.*compile-file-pathname*)

This rest is supposed to return VALUES T T but instead returns:

T
(#P"/home/phoe/Projects/Git/ansi-test/sandbox/compile-file-test-file-5.lsp"
 #P"home:Projects;Git;ansi-test;sandbox;compile-file-test-file-5.lsp.newest")

The spec says:
  The value of *COMPILE-FILE-PATHNAME* must always be a pathname or nil.

So *COMPILE-FILE-PATHNAME* is allowed to be a physical OR a logical
pathname, BUT:
  During a call to COMPILE-FILE, *COMPILE-FILE-PATHNAME* is bound to the
  pathname denoted by the first argument to COMPILE-FILE, merged against the
  defaults; that is, it is bound to (PATHNAME (MERGE-PATHNAMES INPUT-FILE)).

The spec says on MERGE-PATHNAMES:
  MERGE-PATHNAMES returns a logical pathname if and only if its first arcument
  is a logical pathname, or its first argument is a logical pathname
  namestring with an explicit host, or its first argument does not specify a
  host and the default-pathname is a logical pathname.

First argument is #p"compile-file-test-file-5.lsp" and therefore has no host.
The defaults for me are`#P"/home/phoe/Projects/Git/ansi-test/sandbox/"` - a
physical pathname. So the first call MERGE-PATHNAMES properly returns a
physical pathname.

And therefore, inside the compiled file, the value of *COMPILE-FILE-PATHNAME*
is incorrect. We solve this by modifying CCL::%COMPILE-FILE and removing an
explicit but unnecessary (and non-conforming) call to
CCL::BACK-TRANSLATE-PATHNAME that modifies the value that is then bound to
*COMPILE-FILE-PATHNAME* inside the compiled file.
The body of the failing test:

(deftest subtypep.cons.43
  (let* ((n -3.926510009989861d7)
         (t1 '(not (cons float t)))
         (t2 `(or (not (cons (eql 0) (real ,n ,n)))
                  (not (cons t (eql 0))))))
    (multiple-value-bind
     (sub1 good1)
     (subtypep* t1 t2)
     (multiple-value-bind
      (sub2 good2)
      (subtypep* `(not ,t2) `(not ,t1))
      (or (not good1)
          (not good2)
          (and sub1 sub2)
          (and (not sub1) (not sub2))))))
  t)

We are only interested in a part of this test. After simplifying:

(let* ((t1 '(not (cons float t)))
       (t2 `(or (not (cons (eql 0) (real -3.5d0 -3.5d0)))
                (not (cons t (eql 0))))))
  (subtypep t1 t2))

This evaluates to (VALUES NIL T), even though T2 is equivalent to T and
therefore the valid results are (VALUES T T) and (VALUES NIL NIL).

After trying to find a smaller reproducible test case, I have found
that the error is triggered with the following two types:

CCL> (csubtypep (specifier-type
                 '(CONS (NOT FLOAT) T))
                (specifier-type
                 `(OR (CONS (INTEGER 0 0)
                            (OR (REAL * (-3.5D0))
                                (INTEGER * -1)
                                (NOT INTEGER)
                                (REAL (-3.5D0))))
                      (CONS (OR (NOT INTEGER)
                                (INTEGER * -1)
                                (INTEGER 1))
                            T))))
;=> NIL, T

Simplifying again:

CCL> (specifier-type '(OR (REAL * (-3.5D0))
                          (NOT INTEGER)
                          (REAL (-3.5D0))))

The above type is equivalent to T, but CCL fails to infer that fact.

We fix this issue by modifying CCL::SIMPLIFY-UNIONS and introducing
a special case for merging numeric types that are neighboring and have
a common exclusive bound (one lower, one higher) in presence of other types
that might contain that exclusive bound. This causes CCL to correctly
simplify the above type into T.
The test in question is:

(deftest map.error.11
  (let ((type '(or (vector t 5) (vector t 10))))
    (if (subtypep type 'vector)
        (eval `(signals-error (map ',type #'identity '(1 2 3 4 5 6))
                              type-error))
      t))
  t)

This fails because (coerce '(1 2 3 4 5 6) '(or (vector t 5) (vector t 10)))
does not signal an error in safe code. We "fix" this issue by following
SBCL's lead in the matter and signaling an error whenever an OR type is passed
to COERCE and the expected type is a subtype of VECTOR; a proper fix is not in
place due to the amount of work required to perform the typecheck.

This in turn causes ANSI-TEST MAP.48 to break. It expects the following form

(map '(or (vector t 10) (vector t 5)) #'identity '(1 2 3 4 5))

to return #(1 2 3 4 5). We work around this issue by disabling note
:RESULT-TYPE-ELEMENT-TYPE-BY-SUBTYPE in ANSI-TEST.

Signed-off-by: Michał phoe Herda <[email protected]>
This commit removes CCL's implementation of ~F FORMAT control and the CCL
copy of FLONUM-TO-STRING, replacing these with their respective SBCL
counter parts. This was made in order to make all ANSI-TESTs from the
FORMAT.F.* group to pass.

This commit additionally modifies the internals of PRINT-OBJECT method for
floats in order to account for the modified FLONUM-TO-STRING.
CCL incorrectly formatted (format nil "~2f ~2f" 1.1 1.9) - the proper result
is "1.0 2.0" due to the reasoning provided in ANSI-TEST:

    Rationale for FORMAT.F.45: CLHS 22.3.3.1 states that "d is the number of
    digits to print after the decimal point;" AND that the number is printed
    "rounded to d fractional digits".

    If we want to print 1.1 in a field of width 0, then we compute D to be W,
    minus the number of digits to be printed before the decimal point, minus 1
    for the decimal point. So, in this case, D = W - 1 - 1 = 0.

    This means that we must print exactly 0 digits after the decimal point and
    that we must round the number to 0 fractional digits. The latter is doable
    and we round 1.1 to 1.0, BUT the first contradicts the first paragraph of
    22.3.3.1 which states, "arg is printed as a float". "1." is not a float in
    Common Lisp, since it is read as the integer 1 in base 10.

    Therefore, in order to work around this corner case, we explicitly print one
    decimal digit so that the resulting number is recognizable as a float. Since
    the number was rounded to 0 decimal digits, then this digit is 0. This way,
    we get "1." concatenated with "0" that gives us "1.0".
CCL incorrectly printed 0.01 with ~1f and ~0f as ".01", where the correct
value is ".0". This commit fixes this bug.
@xrme
Copy link
Member

xrme commented Jan 7, 2020

Is this still ready to go? Any reason I shouldn't merge it?

@phoe
Copy link
Contributor Author

phoe commented Jan 7, 2020

I guess that it is still good to go - can't see any reason why not.

@phoe
Copy link
Contributor Author

phoe commented Jan 7, 2020

One more thing though, just to be sure - I'd like Clozure/ccl to have Travis configured on it in order to run the ccl-test automated test suite on it, just to be sure we do not break anything.

I will prepare a PR with this, however, it'll require a few steps for the repository administrator (read: not me) to authenticate with Travis.

@kpoeck
Copy link
Contributor

kpoeck commented Jan 18, 2020

Phoe I get both test-errors solved and new test-errors. Did you change some conditions for the tests, e.g. with notes? I defined rt::expected-failures to get this outpupt.

17 unexpected failures: MAP.48, PRINT.SHORT-FLOAT.RANDOM, PRINT.SINGLE-FLOAT.RANDOM, 
   PRINT.DOUBLE-FLOAT.RANDOM, PRINT.LONG-FLOAT.RANDOM, FORMAT.F.46B, 
   FORMATTER.F.46B, FORMAT.E.10, FORMAT.E.11, FORMAT.E.12, FORMAT.E.13, 
   FORMAT.E.14, FORMAT.E.15, FORMAT.E.22, FORMAT.E.23, FORMAT.E.24, FORMAT.E.25.
14 unexpected successes: 
   SYNTAX.SHARP-COLON.ERROR.1, SHIFTF.7, MACROLET.36, MAP.ERROR.11, 
   SUBTYPEP.CONS.43, FORMAT.F.5, FORMAT.F.8, FORMAT.F.45, FORMATTER.F.45, 
   FORMAT.F.46, FORMATTER.F.46, FORMAT.F.47, FORMATTER.F.47, COMPILE-FILE.16.

running the self tests of ccl also produces some messages:

 ================ Test suite failed ================

8 out of 21908 total tests failed: 
   CL-TEST::MAP.48, CL-TEST::PRINT.SHORT-FLOAT.RANDOM, 
   CL-TEST::PRINT.SINGLE-FLOAT.RANDOM, CL-TEST::PRINT.DOUBLE-FLOAT.RANDOM, 
   CL-TEST::PRINT.LONG-FLOAT.RANDOM, CL-TEST::FORMAT.F.5, CL-TEST::FORMAT.F.8, 
   CL-TEST::CCL.BUG\#1186.

@phoe
Copy link
Contributor Author

phoe commented Jan 23, 2020

@kpoeck I remember that I have made some issues on the repository at https://gitlab.common-lisp.net/ansi-test/ansi-test that comment on some of the tests. I am also disabling two of the ansi-test notes:

(disable-note :ansi-spec-problem)
(disable-note :result-type-element-type-by-subtype)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet

3 participants