[clisp-list] Cross-compiling adventure.

[Kaz K. <kky...@gm...>]

Hi all,

I'm cross-compiling CLISP 2.45 on an Intel Linux box, targetting a MIPS
Linux system with the n32 ABI.

I had to patch a few things to get as far as making a lisp.run
executable, linked with all the goodies: sigsegv, ffcall.

(After making lisp.run, the Makefile then bails: it wants to update the
interpreted.mem target, and of course makemake doesn't generate such a
rule, since running lisp.run to make interpreted.mem is impossible under
cross-compiling. It's probably trying that target only because I invoked
make wrong; I gave it no specific target).

It's obvious from the Makefile that that there are some cross-compiling
rules which require a clisp that is compiled to run on the build
machine.  This clisp is to be used as a cross compiler. The rules
require the presence of a special memory image: "clisp -M cross.mem",
and in this image there is supposed to be a function named
CROSS:COMPILE-FILE which is invoked using a -x argument.

Problem is, there isn't any rule to make cross.mem, nor do there appear
to be any materials from which it would be made!

Do share the secret!

Patches:

This is needed because the my 2.6.26 asm-mips/sigcontext.h kernel header
tries to define struct sigcontext, which clashes with what has already
been pulled in from a glibc header.

Index: libsigsegv-2.5/src/fault-linux-mips.h
===================================================================
--- libsigsegv-2.5.orig/src/fault-linux-mips.h  2005-05-20
12:47:21.000000000 -0700
+++ libsigsegv-2.5/src/fault-linux-mips.h       2009-01-15
07:47:10.850946160 -0800
@@ -15,8 +15,6 @@
   along with this program; if not, write to the Free Software Foundation,
   Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.  */

-#include <asm/sigcontext.h>
-
 #define SIGSEGV_FAULT_HANDLER_ARGLIST  int sig, int code, struct
sigcontext *scp
 #define SIGSEGV_FAULT_CONTEXT  scp
 #define SIGSEGV_FAULT_STACKPOINTER  scp->sc_regs[29]


Next are CLISP patches.

Firstly, if the architecture happens to be mips64-*, that doesn't
necessarily mean that the 64 bit ABI is being used. I have a compiler
called mips64-linux-gcc, but I'm using it with the n32 ABI, which is
default. So I hacked an ABI check into makemake.  It's not a perfect
solution, because it relies on inspecting the compiler's default ABI,
which could be overridden by CFLAGS, or maybe the contents of COMPILER
itself.  It would be better to have a configure parameter for selecting
the ABI. All of this has to do with the arithmetic module, arimips.d and
arimips64.d.

In the next hunk, we deal with the bug with comment5. It's being linked
against a gllib/malloc.o, which is a MIPS object file. We need comment5
to be a build machine tool.  Rather than compile malloc.c again to a
build machine .o,  I just patched things so that the .c file is passed
to the compiler when building comment5. It's a trivial small thing; no
harm in compiling it a few times over again.

The next hunk addresses the problem that I got unresolved symbols,
because the final link didn't include the ffcall libraries. In makemake,
a FFCALL_LIBS variable is defined, but not used!

The last change is a hack to get around the problem that the arithmetic
assembly file arimips.s was being compiled with /bin/as. This is because
TSYSOS is not simply "linux", but "mips-gnu-linux-none" or something
like that. This comes in as a parameter when configure runs makemake.  I
just blew away that logic so it unconditionally uses the cross GCC.

===================================================================
--- clisp-2.45.orig/src/makemake.in     2008-05-08 18:22:01.000000000 -0700
+++ clisp-2.45/src/makemake.in  2009-01-15 17:39:58.309870528 -0800
@@ -1048,7 +1048,22 @@
  cpu=mips
 fi
 if [ "$host_cpu" = mips64 ] ; then
+  # cpu value depends on ABI, which we get from the compiler if it's
GCC cpu=mips64
+  case "$COMPILER" in
+  *gcc*)
+    for config in `$COMPILER -v 2>&1` ; do
+      case "$config" in
+      --with-abi=*32 )
+       cpu=mips
+       ;;
+      --with-abi=*64 )
+       cpu=mips64
+       ;;
+      esac
+    done
+    ;;
+  esac
 fi
 if [ "$host_cpu" = i386 -o "$host_cpu" = i486 -o "$host_cpu" = i586
-o "$host_cpu" = i686 -o "$host_cpu" = i786 -o $TSYS = i386 -o $TSYS =
i486 -o $TSYS = i586 -o $TSYS = i686 -o $TSYS = i786 -o $TOS = win32 ]
; then
  cpu=i386
@@ -1501,7 +1516,8 @@
 if [ ${HAVE_MALLOC_POSIX} = 0 -o ${CROSS} = true ]; then
  # gnulib/stdlib.h will define malloc to rpl_malloc, so we need this
  GLLIB_O=gllib/malloc${TOBJ}; GLLIB=gllib/malloc
-  EXTRAOBJ=" gllib/malloc${TOBJ}"
+  EXTRAOBJ=" ${SRCDIR}/gllib/malloc.c"
+  CPPFLAGS="${CPPFLAGS} -I."
 else GLLIB_O=""; GLLIB=""; EXTRAOBJ="";
 fi
 for f in localcharset uniname/uniname uniwidth/width; do
@@ -1915,7 +1931,7 @@
  echol "XCFLAGS = ${XCFLAGS}"
  echol "XCPP = ${XCPP}"
  echol "XCLFLAGS = ${XCLFLAGS}"
-  echol "XLIBS = ${LIBS}"
+  echol "XLIBS = ${LIBS} ${FFCALL_LIBS}"
  XCC="\$(XCC)"
  XCPPFLAGS="\$(XCPPFLAGS)"
  XCFLAGS="\$(XCFLAGS)"
@@ -2746,16 +2762,11 @@
      # Only the native as groks the .SHORTDATA statements in arihppa.d
      echotab "${XCC} ${XCPPFLAGS} ${XCFLAGS} -c ${f}.s ||
/usr/ccs/bin/as ${f}.s -o ${f}${TOBJ} || /bin/as ${f}.s -o
${f}${TOBJ}"
    else
-      if [ $f = arimips -a "$TSYSOS" != linux ] ; then
-        # Avoid problem during incremental linking of lisp.o.
-        echotab "/bin/as -KPIC ${f}.s -o ${f}${TOBJ} ||
/usr/ccs/bin/as -KPIC ${f}.s -o ${f}${TOBJ}"
-      else
        if [ $XCC_GCC = true ] ; then
          echotab "${XCC} ${XCPPFLAGS} ${XCFLAGS} -x assembler -c ${f}.s"
        else
          echotab "${XCC} ${XCPPFLAGS} ${XCFLAGS} -c ${f}.s"
        fi
-      fi
    fi
    echol
  fi