Problems with differing *READ-BASE* values for COMPILE-FILE/LOAD

[Mark Evenson]

Eric Timmons reports

For file-2.test

(defun test () 10)

exercised with

(in-package :cl-user)

(let ((*read-base* 2))
  (compile-file "file-2.lisp"))

(load "file-2.abcl")

(format t "Got ~D, expected 2~%" (test))

(fmakunbound 'test)

(let ((*read-base* 9))
  (compile-file "file-2.lisp"))

(load "file-2.abcl")

(format t "Got ~D, expected 9~%" (test))

[Mark Evenson]

Fixed in abcl-1.8.1-dev with recent work on the loader.

[Eric Timmons]

I can still reproduce on commit 9c0a9693ea933309e503266255edcde5cc76257c. Here's a modified test procedure that gets rid of the *read-base* nonsense.

(defvar *file-name* "file-2")

(defun write-test-file (num)
  (let ((pn (merge-pathnames *file-name* (make-pathname :type "lisp"))))
    (with-open-file (f pn
                       :direction :output
                       :if-exists :supersede)
      (prin1 `(defun test () ,num) f))
    pn))

(load (compile-file (write-test-file 10)))

(assert (= (test) 10) nil "Got ~D, expected 10" (test))

(delete-file (merge-pathnames *file-name* (make-pathname :type "abcl")))

(load (compile-file (write-test-file 11)))

(assert (= (test) 11) nil "Got ~D, expected 11" (test))

The extra weird thing is that if you spawn a new ABCL process and LOAD file-2.abcl, evaluating (TEST) produces the correct result (11), so the COMPILE-FILE works. It's acting like the first LOAD of an .abcl file caches the file handle, doesn't close it, and reuses it on subsequent LOADs.

Bisecting suggests this appeared in one of:

# only skipped commits left to test
# possible first bad commit: [2a1e20db8cc46edb277d3415d9e6dbc0f65fa671] Ensure that we behave like previous implementation
# possible first bad commit: [6824d00bb83a8032aa7716a211719118e6404614] Make Pathname.init() static method
# possible first bad commit: [bd3e7de6aa87d1fdc6817ded0562c792ba4db247] Starting to hook jar/uri mechanisms up
# possible first bad commit: [fc63b71217430274904eced6c333091d13980b4a] INCOMPLETE Hook up implementation of PathnameJar and PathnameURL

EDIT:

CL-USER> (lisp-implementation-version)
"1.8.1-dev"
"OpenJDK_64-Bit_Server_VM-AdoptOpenJDK-1.8.0_272-b10"
"amd64-Linux-5.4.73-gentoo"

[Mark Evenson]

Replying to Eric Timmons:

The extra weird thing is that if you spawn a new ABCL process and LOAD file-2.abcl, evaluating (TEST) produces the correct result (11), so the COMPILE-FILE works. It's acting like the first LOAD of an .abcl file caches the file handle, doesn't close it, and reuses it on subsequent LOADs.

The failure to re-load FASLs is an artifact of the a bug with the zip cache that I have been working on fixing. Unfortunately, I didn't note this as a problem publicly, so apologize for not having all the information available for you.

I think the zip cache is now fixed with <​https://github.com/armedbear/abcl/pull/357> which is more-or-less ready to be merged. But first, I am transitioning to Github Actions for CI away from Travis due to their recent limitations on macOS runners for open source projects which I easily blew away in a day's worth of CI.

[Mark Evenson]

Finished adding Github Actions for CI, and merged the relevant patches.

With <https://abcl.org/trac/changeset/15494> your test code now works.

Closing again, but please re-open with more comments if this is still broken somehow.