Re: [Gcl-devel] Re: possible GCL/Windows compiler bug

[Matt Kaufmann]

P.S. Jared just did an experiment for me that attempts to cause the same
problem using functions defined by the user, in place of ACL2 source function
match-clause and others.  But the break didn't occur, which is pretty
interesting.  Apparently it may be that conditions have to somehow be just
right for this bug to show up.

I'll probably try to get the same problem to occur tonight on my dual-boot
machine at home by installing ACL2 on the Windows side (Jared reminded me that
he has an installer that makes this easy).

-- Matt
   Date: 12 Oct 2004 14:44:59 -0500
   From: Matt Kaufmann <address@hidden>
   CC: address@hidden, address@hidden

   Here is some additional information that I think will affect how you want to
   proceed.

   First of all, the offending function, ACL2_*1*_ACL2::MATCH-CLAUSE, isn't
   defined in basis.lisp (unlike ACL2::MATCH-CLAUSE, which _is_ defined in
   basis.lisp).  Rather, the definition, def, of ACL2_*1*_ACL2::MATCH-CLAUSE is
   generated and compiled on the fly during the build, with (eval def) and
   (eval `(compile ',(cadr def))).

   So, loading basis.o presumably won't help.  Instead, I hacked the build
   procedure so that for ACL2_*1*_ACL2::MATCH-CLAUSE, we write out a file
   my-debug.lisp:

   
===============================================================================
   (in-package "ACL2")

   (DEFUN ACL2_*1*_ACL2::MATCH-CLAUSE (X PAT FORMS)
     (COND
       ((F-GET-GLOBAL 'SAFE-MODE *THE-LIVE-STATE*)
        (RETURN-FROM ACL2_*1*_ACL2::MATCH-CLAUSE
          (MV-LET (TESTS BINDINGS)
                  (ACL2_*1*_ACL2::MATCH-TESTS-AND-BINDINGS X PAT NIL NIL)
                  (LIST (IF (ACL2_*1*_COMMON-LISP::NULL TESTS) T
                            (CONS 'AND
                                  (ACL2_*1*_COMMON-LISP::REVERSE TESTS)))
                        (CONS 'LET
                              (CONS (ACL2_*1*_COMMON-LISP::REVERSE
                                        BINDINGS)
                                    FORMS)))))))
     (MATCH-CLAUSE X PAT FORMS))
   
===============================================================================

   Instead of (eval `(compile ',(cadr def))), the hacked build code does the
   following during the build:

   (compile-file "my-debug.lisp" :c-file t :h-file t)
   (load "my-debug")

   I figured that this is what you'd need in order to carry out your plan.

   Jared Davis graciously ran this experiment.  Unfortunately, with that small
   change in the build procedure the problem goes away.  (I suppose we could
   change ACL2 to do this for all functions, but it seems unfortunate to do all
   that unnecessary file io, and I wonder if that would mask some other issue.)

   So, how would you like to proceed?

   Thanks --
   -- Matt
      Cc: address@hidden, address@hidden
      From: Camm Maguire <address@hidden>
      Date: 12 Oct 2004 11:59:34 -0400
      User-Agent: Gnus/5.09 (Gnus v5.9.0) Emacs/21.2
      Content-Type: text/plain; charset=us-ascii
      X-SpamAssassin-Status: No, hits=-2.6 required=5.0
      X-UTCS-Spam-Status: No, hits=-322 required=180

      Greetings!

      Matt Kaufmann <address@hidden> writes:

      > Hi --
      > 
      > Thanks very much for the quick reply!  I have some questions.
      > 
      > When I tried gdb on ACL2/linux saved_acl2.gcl (built with GCL 2.6.5), 
and executed
      > 
      >   gdb linux-gcl-saved_acl2.gcl
      > 
      > then I got the following unfortunate result (where I edited out the 
pathname):
      > 
      >   GNU gdb 5.3
      >   Copyright 2002 Free Software Foundation, Inc.
      >   GDB is free software, covered by the GNU General Public License, and 
you are
      >   welcome to change it and/or distribute copies of it under certain 
conditions.
      >   Type "show copying" to see the conditions.
      >   There is absolutely no warranty for GDB.  Type "show warranty" for 
details.
      >   This GDB was configured as "i686-pc-linux-gnu"...(no debugging 
symbols found)...
      >   (gdb) b fasload
      >   Function "fasload" not defined.
      >   (gdb) r
      >   Starting program: .../linux-gcl-saved_acl2.gcl 
      >   (no debugging symbols found)...(no debugging symbols found)...(no 
debugging symbols found)...
      >   Program received signal SIGSEGV, Segmentation fault.
      >   0x400c5d45 in memset () from /lib/libc.so.6
      >   (gdb) 
      > 

      Forgot you have sgc on.  Do 'handle SIGSEGV nostop noprint' here and
      continue with 'c'

      > Is it necessary to build ACL2 with some special settings somehow in 
order to
      > get debug info?
      > 

      In general, yes, but we can save time by checking quickly if we can
      pinpoint the error in the mentioned function.  In general, we build
      gcl with --enable-debug, but you can also get part of the way by
      setting compiler::*c-debug* to t.

      BTW, looking at the C source, I noticed native gcl support for set-mv
      and mv-ref.  Nice to see these two projects so closely linked.

      If/when you rebuild, please also do so with compiler::*default-c-file*
      set to t so we can keep the generated C source just in case.  I
      strongly doubt it is any different than what I can generate under
      Linux. 

      > In 4), how do you submit a Lisp LOAD command inside gdb?  Also, in 2), 
how do

      Once you type 'r', you will have a lisp prompt.

      > we arrange that r will run the particular commands that triggered the 
break?
      > 

      Starting in 3) issue commands to lisp as normal to trigger the
      error. 'r' just starts acl2.

      > General issue:  Recall that when we re-compile the match-clause 
function, the
      > error goes away.  Doesn't that suggest that your approach won't trigger 
the
      > error in 8)?
      > 

      Missed this somehow.  If it does not trigger the error, then trigger
      it from gdb running acl2/lisp as you know how, send the backtrace,
      send objdump -d basis.o, find the address (printed) where basis.o was
      loaded, find out the address where the fault occurs, and 'p/x *(char
      *)<basis.o load address>@1024', 'p/x *((char *)<basis.o load
      address>+1024)@1024', etc. until you print out the address of the
      fault.  You may have difficulty locating the fault address with sgc
      on.  See if you can retain the error with sgc turned off.  If not,
      then 'b sgbc.c:1626' and 'cond 1 fault_count > 300' (assuming the
      breakpoint just created was numbered 1, and 'b segmentation_catcher'.
      You can then see the fault address in the backtrace (bt) printed under
      gdb. 

      Take care,

      > Thanks --
      > -- Matt
      >    Cc: address@hidden, "Mike Thomas" <address@hidden>,
      >            address@hidden
      >    From: Camm Maguire <address@hidden>
      >    Date: 12 Oct 2004 10:40:23 -0400
      >    User-Agent: Gnus/5.09 (Gnus v5.9.0) Emacs/21.2
      >    Content-Type: text/plain; charset=us-ascii
      >    X-SpamAssassin-Status: No, hits=-2.6 required=5.0
      >    X-UTCS-Spam-Status: No, hits=-332 required=180
      > 
      >    Greetings!
      > 
      >    OK, my suspicion lies with a difficulty in the windows relocation
      >    code, but I'm not yet certain.  Here is how to proceed:
      > 
      >    0) run saved_acl2 under gdb 
      > 
      >    1) b fasload
      > 
      >    2) r
      > 
      >    3) Put your match-clause function into a separate file, compile with
      >       compiler::*c-debug* set to t
      > 
      >    4) (load "your_file.o")
      > 
      >    5) gdb will break -- type 'finish'
      > 
      >    6) Still at gdb prompt, look at start address printed by gcl to
      >       screen.  Type 'add-symbol-file <your_file.o> <address>
      > 
      >    7) c
      > 
      >    8) trigger the error
      > 
      >    9) gdb will stop, print a backtrace with bt
      > 
      >    10) In gdb, type 'p/x *(char *)<address>@1024'
      > 
      >    11) In gdb, type 'shell', and then 'objdump -d your_file.o'
      > 
      >    12) Send me the results
      > 
      >    Take care,
      > 
      > 
      >    Matt Kaufmann <address@hidden> writes:
      > 
      >    > [Resending -- I had a typo in the CC field that may have prevented 
delivery.]
      >    > 
      >    > Hi --
      >    > 
      >    > Help!?  Sorry to bother you with this email, but I've gone about 
as far as I
      >    > know how with the problem described below (and I'll spare you the 
dead ends),
      >    > which I kind of suspect is a problem with GCL/Windows, but might 
instead be a
      >    > problem with ACL2.  This is a rather long email; please feel free 
to ask for
      >    > any clarification.
      >    > 
      >    > There appears to be a problem either with GCL on Windows or with 
ACL2.  I'll
      >    > describe the symptom below.  This symptom doesn't occur on Linux 
or Sun/Solaris
      >    > for GCL, Allegro CL, CMUCL, or CLISP, and I also haven't seen it 
on Linux with
      >    > Lispworks or on a Macintosh with OpenMCL; I've only seen it on 
GCL/Windows.
      >    > But I realize that it still could be a subtle problem with ACL2, 
so I think we
      >    > need to wait on the ACL2 release until we determine whether or not 
it's a
      >    > GCL/Windows problem.
      >    > 
      >    > Jared Davis was kind enough to submit the commands below (at the 
end of this
      >    > email) while standing in directory books/misc/ of the ACL2 
distribution, after
      >    > building ACL2 on GCL/Windows 2.6.5.  The result is a hard Lisp 
error if you do
      >    > *NOT* submit the compile form below: for a transcript, see
      >    > http://www.cs.utexas.edu/users/jared/test.log5 (or equivalently, 
on the UTCS
      >    > file system, /u/www/users/jared/test.log5).  But *with* the 
compile form, the
      >    > problem goes away (see test.log4 in the same directory).  We 
verified that the
      >    > definition being compiled is exactly the same as the one compiled 
when building
      >    > ACL2.  (I'll explain how if you're interested -- we could insert a 
call of
      >    > disassemble during the build if you think that would be helpful.)
      >    > 
      >    > You can see the result of :bt and :ihs on a failed run, where some 
source
      >    > functions are run interpreted (but this doesn't avoid the error 
since the
      >    > offending function ACL2_*1*_ACL2::MATCH-CLAUSE is still run 
compiled), in
      >    > test.log in that same directory.  In a moment I'll forward you a 
related log
      >    > with that info (and also the result of :bl), in case you prefer to 
see it by
      >    > email.
      >    > 
      >    > By the way, all of the failures Jared came across during the 
regression run
      >    > were during macroexpansion of ACL2 macro case-match, which calls
      >    > match-clause-list, which calls match-clause -- actually the ACL2 
macroexpansion
      >    > mechanism causes a call of ACL2_*1*_ACL2::MATCH-CLAUSE-LIST, which 
calls
      >    > ACL2_*1*_ACL2::MATCH-CLAUSE.
      >    > 
      >    > Also by the way, even if you leave off the compile form below but 
you add
      >    > (si::use-fast-links nil), the problem goes away.  That seems odd 
to me so I
      >    > thought I should mention it.
      >    > 
      >    > Here are the commands after starting up ACL2.  Again, omit the 
compile form to
      >    > see the error -- even though the compile form should be a no-op!
      >    > 
      >    > (rebuild "defpun.lisp" 'arbitrary-tail-recursive-encap)
      >    > :q
      >    > (compile
      >    >  (DEFUN ACL2_*1*_ACL2::MATCH-CLAUSE (X PAT FORMS)
      >    >    (COND
      >    >     ((F-GET-GLOBAL 'SAFE-MODE *THE-LIVE-STATE*)
      >    >      (RETURN-FROM
      >    >       ACL2_*1*_ACL2::MATCH-CLAUSE
      >    >       (MV-LET (TESTS BINDINGS)
      >    >               (ACL2_*1*_ACL2::MATCH-TESTS-AND-BINDINGS X PAT NIL 
NIL)
      >    >               (LIST (IF (ACL2_*1*_COMMON-LISP::NULL TESTS) T
      >    >                         (CONS 'AND
      >    >                               (ACL2_*1*_COMMON-LISP::REVERSE 
TESTS)))
      >    >                     (CONS 'LET
      >    >                           (CONS (ACL2_*1*_COMMON-LISP::REVERSE
      >    >                                  BINDINGS)
      >    >                                 FORMS)))))))
      >    >    (MATCH-CLAUSE X PAT FORMS)))
      >    > (lp)
      >    > (defun remove-xargs-domain-and-measure (dcl)
      >    >   (case-match dcl
      >    >     (('declare ('xargs ':domain dom-expr
      >    >                        ':measure measure-expr
      >    >                        . rest))
      >    >      (mv nil dom-expr measure-expr rest))
      >    >     (('declare ('xargs ':gdomain dom-expr
      >    >                        ':measure measure-expr
      >    >                        . rest))
      >    >      (mv t dom-expr measure-expr rest))
      >    >     (& (mv nil nil 0 nil))))
      >    > 
      >    > Thanks much --
      >    > -- Matt
      >    > 
      >    > 
      >    > 
      > 
      >    -- 
      >    Camm Maguire                                         address@hidden
      >    
==========================================================================
      >    "The earth is but one country, and mankind its citizens."  --  
Baha'u'llah
      > 
      > 
      > _______________________________________________
      > Gcl-devel mailing list
      > address@hidden
      > http://lists.gnu.org/mailman/listinfo/gcl-devel
      > 
      > 
      > 

      -- 
      Camm Maguire                                              address@hidden
      ==========================================================================
      "The earth is but one country, and mankind its citizens."  --  Baha'u'llah