Re: New gdb 31 & 64 bit patches for S/390

[Eli Zaretskii <eliz@is.elta.co.il>]

On Thu, 5 Jul 2001, Andrew Cagney wrote:

> > What about using the __attribute__(packed) gcc extension.
> > & add a
> > #ifndef gcc
> > define __attribute__
> > #endif
> 
> No.  So far GDB has managed to avoid a dependency on GCCoteric features, 
> I don't see any reason to change this.
> 
> With regard to the other target specific structures, I suggested moving 
> them to s390-nat.c since (I think) only that file would be using them 
> (?correct). s390-nat.c is very host=target specific - it needs to 
> correctly unpack the data returned from ptrace/procfs.  However, even 
> there, the __attribute__(packed) should be removed.

If taken at face value, IMHO this is too harsh to the developers.

I agree that compiler-specific extensions should be kept at the bare
minimum, but why are you opposed to __attribute__((packed)) in native
files?  Some functionality is impossible to get right without that.
How else can I define a struct which fits some external OS data
structure which is not under my control?  The only way I know of is to
use a char array with ugly, hand-computed, error-prone offsets into it
and lots of type casts to fetch and store data there.  Do we really
want that kind of ugliness in GDB?

For example, here's a definition of an ia32 segment descriptor:

    struct seg_descr {
      unsigned short limit0          __attribute__((packed));
      unsigned short base0           __attribute__((packed));
      unsigned char  base1           __attribute__((packed));
      unsigned       stype:5         __attribute__((packed));
      unsigned       dpl:2           __attribute__((packed));
      unsigned       present:1       __attribute__((packed));
      unsigned       limit1:4        __attribute__((packed));
      unsigned       available:1     __attribute__((packed));
      unsigned       dummy:1         __attribute__((packed));
      unsigned       bit32:1         __attribute__((packed));
      unsigned       page_granular:1 __attribute__((packed));
      unsigned char  base2           __attribute__((packed));
    };

How do I define something like that without packing, and make sure it
works with any version of GCC, past and future?

It's clear that something like this can only be put into a native file
which is only compiled by GCC.  But given that those constraints are
satisfied, what's the problem with having this in GDB?