Re: Patch for bug 567 (sparc-sun-solaris2.8)

[Duncan Roe <duncanr@optimation.com.au>]

Hi Andrew,

I posted the debug info the other day, as you requested.

What happens now?

Cheers ... Duncan.

On Fri, Nov 15, 2002 at 07:34:58PM -0500, Andrew Cagney wrote:
> > Hi,
> > 
> > Bug 567 complains that when you build gdb for sparc-sun-solaris2.8 as a 64-bit
> > program, then when you use it to debug a 32-bit program things happen like:
> > 
> >    (gdb) p d1
> >    Cannot access memory at address 0xffbef7a0
> >    (gdb) x/fg &d1
> >    0xffbef7a0:     3.2999999999999998
> > 
> > (Actually 567 shows the wrong value being printed to x/fg, but this seems to be
> > fixed at gcc-3.2 / gdb 5.2.1).
> > 
> > It turns out that you can't "p" any kind of variable: the problem isn't limited
> > to double.
> > 
> > The problem seemed to me that 32-bit negative stack offsets became +ve 64-bit
> > quantities. This would give a 33-bit address (the wanted address with an extra
> > "1" on the left).
> > 
> > Ideally, I think gdb should just mask off the extra bit when debugging 32-bit
> > code. But I couldn't figure out how to do that, so the attached patch
> > sign-extends symbol values as they are read in.
> 
>  From the point of view of GDB's core, the patch below may be closer to 
> correct than you think (don't know how well it fits into the definition 
> of the debug info - debug maintainer problem :-).
> 
> GDB converts all external (debug info, et.al.) addresses into a 
> canonical form.  That form won't involve masking but can, on ocasions, 
> involve sign extension.  This is so that GDB can correctly debug a 32 
> bit ABI on a 64 bit target.  In such a situtation, while a pointer might 
> be 32 bits, registers and the address space would be the full 64 bits, 
> gdb extending everything out to the size of CORE_ADDR.
> 
> This is what makes it possible for GDB to debug an o32 ABI on a MIPS 64 
> platform (eg IRIX 6.5).
> 
> BTW, can you post the corresponding debug info?
> 
> Andrew
> 
> 
> > GCC 3.2 configuration:
> > 
> >  /tmp/gcc-3.2/configure --prefix=/usr/local/gcc-3.2
> > 
> > GDB 5.2.1 configuration
> > 
> >  PATH=/usr/local/gcc-3.2/bin:$PATH
> >  export PATH
> >  CC="gcc -m64"
> >  export CC
> >  CFLAGS="-g -O2"
> >  export CFLAGS
> >  ./configure --prefix=/usr/local/solaris2.8_64
> > 
> > You need to put "-m64" in CC rather than CFLAGS else "make install" fails when
> > trying to build "chew". This has the unfortunate side-effect that "make check"
> > also uses "gcc -m64", i.e. it never exercises "gcc" which would build a 32-bit
> > program.
> > 
> > Cheers ... Duncan.
> > 
> > 
> > 
> > diff -r -u gdb-5.2.1.bu/gdb/dbxread.c gdb-5.2.1/gdb/dbxread.c
> > --- gdb-5.2.1.bu/gdb/dbxread.c	Fri Apr  5 08:33:49 2002
> > +++ gdb-5.2.1/gdb/dbxread.c	Thu Nov 14 16:23:40 2002
> > @@ -2591,6 +2591,27 @@
> >  	fill_symbuf (abfd);
> >        bufp = &symbuf[symbuf_idx++];
> >        INTERNALIZE_SYMBOL (nlist, bufp, abfd);
> > +
> > +      /* ----------------------------------------------- */
> > +      /* Horrible fix for when gdb is built with "-m64"  */
> > +      /* (sparc-sun-solaris2.8):                         */
> > +      /* sign-extend the 32-bit result in nlist.n_value. */
> > +      /*                                                 */
> > +      /* This fixes the testcase in bug 567,             */
> > +      /* in that you can "p d1" successfully.            */
> > +      /* Actually you can "p" *anything*                 */
> > +      /* (previously, you couldn't).                     */
> > +      /*                                                 */
> > +      /* I expect this will break something else,        */
> > +      /* we'll just have to wait to see what.            */
> > +      /*                                                 */
> > +      /* The proper fix is for gdb to know that it's     */
> > +      /* working on a 32-bit program and                 */
> > +      /* truncate addresses to 32 bits before using them */
> > +      /* ----------------------------------------------- */
> > +      
> > +      nlist.n_value = (long)(int)nlist.n_value;
> > +      
> >        OBJSTAT (objfile, n_stabs++);
> >  
> >        type = bfd_h_get_8 (abfd, bufp->e_type);
> > 
> > 
> > 
> > /*
> >  * From GDB bug report 567
> > gcc -g -m64 -o double double.c
> > gcc -g -o double double.c
> >  *
> >  * The report says to breakpoint on the printf line
> >  * & print the value of d1
> >  */
> > extern int printf(const char *, ...);
> > 
> >     int
> >     main()
> >     {
> >         double d1;
> > 
> >         d1 = 3.3;
> >         printf("d1 = %f\n", d1);
> > 
> >         return 0;
> >     }
> 
> 
>