save-lisp-and-die on sbcl.core breaks some tests

[Sébastien Villemot]

In the Debian package, we do the following at build time in order to set the source location to the system-wide one:

HOME=$(pwd) src/runtime/sbcl --core output/sbcl.core --no-sysinit --no-userinit \
             --eval '(sb-ext:set-sbcl-source-location "/usr/share/sbcl-source/")' \
             --eval '(sb-ext:save-lisp-and-die "output/sbcl.core")'

Then we run the testsuite, and since 2.2.2 it fails in the “pure” tests with:

Unhandled SIMPLE-ERROR in thread #<SB-THREAD:THREAD "main thread" RUNNING
                                    {1001810073}>:
  Symbol value differs: SB-IMPL::**FINALIZER-STORE**

Backtrace for: #<SB-THREAD:THREAD "main thread" RUNNING {1001810073}>
0: (SB-DEBUG::DEBUGGER-DISABLED-HOOK #<SIMPLE-ERROR "Symbol value differs: ~S" {10048C53F3}> #<unused argument> :QUIT T)
1: (SB-DEBUG::RUN-HOOK *INVOKE-DEBUGGER-HOOK* #<SIMPLE-ERROR "Symbol value differs: ~S" {10048C53F3}>)
2: (INVOKE-DEBUGGER #<SIMPLE-ERROR "Symbol value differs: ~S" {10048C53F3}>)
3: (ERROR "Symbol value differs: ~S" SB-IMPL::**FINALIZER-STORE**)
4: (RUN-TESTS::COMPARE-SYMBOL-VALUES ((SB-ALIEN::*ALIEN-CALLABLES* . #<HASH-TABLE :TEST EQ :COUNT 0 {10009E94C3}>) (SB-ALIEN::*ALIEN-FUN-TYPE-VARARGS-DEFAULT* . :UNSPECIFIED) (SB-ALIEN::*ALIEN-CALLBACK-TRAMPOLINES* . #()) (SB-ALIEN::*SHARED-OBJECTS-LOCK* . #<SB-THREAD:MUTEX "shared object list lock" (free)>) (SB-ALIEN::*ALIEN-TYPE-CLASSES* . #<HASH-TABLE :TEST EQL :COUNT 16 {10000196A3}>) (SB-ALIEN::*ALIEN-CALLBACK-INFO*) (SB-ALIEN::*ALIEN-CALLBACKS* . #<HASH-TABLE :TEST EQUAL :COUNT 0 {10009E9303}>) (SB-ALIEN::*ALIEN-CALLBACK-WRAPPERS* . #<HASH-TABLE :TEST EQUAL :COUNT 0 {10009E9403}>) (SB-ALIEN::*ALIEN-TYPE-MATCHES*) (SB-ALIEN::*NEW-AUXILIARY-TYPES*) (SB-REGALLOC:*PACK-OPTIMIZE-SAVES* . T) (SB-REGALLOC:*PACK-ASSIGN-COSTS* . T) ...))
5: (RUN-TESTS::PURE-RUNNER (#P"/tmp/sbcl-2.2.2/tests/allocator.pure.lisp" #P"/tmp/sbcl-2.2.2/tests/arith-2.pure.lisp" #P"/tmp/sbcl-2.2.2/tests/arith-slow.pure.lisp" #P"/tmp/sbcl-2.2.2/tests/arith.pure.lisp" #P"/tmp/sbcl-2.2.2/tests/array.pure.lisp" #P"/tmp/sbcl-2.2.2/tests/assembler.pure.lisp" #P"/tmp/sbcl-2.2.2/tests/avltree.pure.lisp" #P"/tmp/sbcl-2.2.2/tests/backq.pure.lisp" #P"/tmp/sbcl-2.2.2/tests/bad-code.pure.lisp" #P"/tmp/sbcl-2.2.2/tests/bit-bash.pure.lisp" #P"/tmp/sbcl-2.2.2/tests/bivalent-stream.pure.lisp" #P"/tmp/sbcl-2.2.2/tests/bug-1072739.pure.lisp" ...) RUN-TESTS::LOAD-TEST #<SB-SYS:FD-STREAM for "file /tmp/sbcl-2.2.2/tests/test.log" {1003875FA3}>)
6: (RUN-TESTS::RUN-ALL)
7: (SB-INT:SIMPLE-EVAL-IN-LEXENV (RUN-TESTS::RUN-ALL) #<NULL-LEXENV>)
8: (EVAL (RUN-TESTS::RUN-ALL))
9: (SB-IMPL::PROCESS-EVAL/LOAD-OPTIONS ((:EVAL . "(setf sb-ext:*evaluator-mode* :compile)") (:EVAL . "(with-compilation-unit () (load \"run-tests.lisp\"))") (:EVAL . "(run-tests::run-all)")))
10: (SB-IMPL::TOPLEVEL-INIT)
11: ((FLET SB-UNIX::BODY :IN SB-IMPL::START-LISP))
12: ((FLET "WITHOUT-INTERRUPTS-BODY-3" :IN SB-IMPL::START-LISP))
13: (SB-IMPL::START-LISP)

I have verified that the crash of the testsuite is related to the save-lisp-and-die command above.

Also, the problem was not present in 2.1.11 (so it was introduced in the 2.2 series, I did not check in which version exactly).

Do you consider this as a bug? Or should I change the procedure for setting the source location, and if yes, how?

Thanks for your work,

S. Villemot

[Kambiz Darabi (darabi)]

I tried a build without

--eval '(sb-ext:set-sbcl-source-location "/usr/share/sbcl-source/")'

and get the same result.

Reducing the number of tests to these:

- compiler.pure.lisp
- compound-cons.pure.lisp
- condition-2.pure.lisp

triggers the error reproducibly. Running the last two does not lead
to an error.

[Christophe Rhodes (csr21-cantab)]

I'm not sure, but I think that a change in value of **FINALIZER-STORE**
when running "pure" tests is not completely unexpected; it's possible
that the consistency tester for running the tests (which is what this
is) is being too strong in its assertions.

The finalizer store can run out of room, a new one allocated and
assigned (copying old values across). I think it's a bit surprising
that that would happen to a freshly-saved core differently from the core
saved from the build process, but...

[Douglas Katzman (dougk)]

I'll add finalizer-store to the list of ignorable symbols.

[Kambiz Darabi (darabi)]

This is reproducible with the sbcl-2.2.2 tarball with these steps:

sh make.sh --fancy
src/runtime/sbcl --core output/sbcl.core --no-sysinit --no-userinit --eval '(sb-ext:save-lisp-and-die "output/sbcl.core")'
cd tests
./run-tests.sh compiler.pure.lisp compound-cons.pure.lisp condition-2.pure.lisp

on two different systems

$ uname -a
Linux myserver 4.15.0-96-generic #97-Ubuntu SMP Wed Apr 1 03:25:46 UTC 2020 x86_64 x86_64 x86_64 GNU/Linux

$ cat /etc/lsb-release
DISTRIB_ID=Ubuntu
DISTRIB_RELEASE=18.04
DISTRIB_CODENAME=bionic
DISTRIB_DESCRIPTION="Ubuntu 18.04.5 LTS"

and

$ uname -a
Linux mylaptop 5.4.0-104-generic #118-Ubuntu SMP Wed Mar 2 19:02:41 UTC 2022 x86_64 x86_64 x86_64 GNU/Linux

$ cat /etc/lsb-release
DISTRIB_ID=Ubuntu
DISTRIB_RELEASE=20.04
DISTRIB_CODENAME=focal
DISTRIB_DESCRIPTION="Ubuntu 20.04.4 LTS"

[Kambiz Darabi (darabi)]

As discussed on IRC [1], the failure is due to too strict consistency checks for side effects during the 'pure' tests.

I will add a patch to the Debian packaging to ignore the modified SB-IMPL::**FINALIZER-STORE**

[1] https://irclog.tymoon.eu/libera/%23sbcl?around=1648544051#1648544051

[Christophe Rhodes (csr21-cantab)]

I believe this was fixed in sbcl-2.2.4