DWARF-2 and constructors/destructors

DWARF-2 and constructors/destructors

[Daniel Jacobowitz]

Suppose you want to call a (non-virtual) method in C++, from something
with Dwarf-2 info.  The only way to get the mangled name from the debug
info is DW_AT_MIPS_linkage_name.  This isn't, of course, present for
constructors/destructors, since the entry in the class definition is
for the abstract version.

Is it reasonable for the debugger to have to mangle this itself?  The
constructor arguments can be arbitrarily complex.  Should there be
references in the debug information to the base and complete
constructors anywhere?

-- 
Daniel Jacobowitz                           Carnegie Mellon University
MontaVista Software                         Debian GNU/Linux Developer

Re: DWARF-2 and constructors/destructors

[Daniel Jacobowitz]

On Wed, Dec 05, 2001 at 10:52:46PM -0500, Daniel Jacobowitz wrote:
> Suppose you want to call a (non-virtual) method in C++, from something
> with Dwarf-2 info.  The only way to get the mangled name from the debug
> info is DW_AT_MIPS_linkage_name.  This isn't, of course, present for
> constructors/destructors, since the entry in the class definition is
> for the abstract version.
> 
> Is it reasonable for the debugger to have to mangle this itself?  The
> constructor arguments can be arbitrarily complex.  Should there be
> references in the debug information to the base and complete
> constructors anywhere?

And I see, in a rather memorable rant, Daniel Berlin saying that
DW_AT_MIPS_linkage_name is "a hack" (certainly), "not necessary, and
we're trying to make it go away anyway."  Dan, would you mind
enlightening me on how to cope without it without having a full mangler
for every supported C++ ABI attached to the Dwarf-2 reader?

As far as I'm concerned, GDB has no business ever mangling a single
string.  Debug info should provide everything that we need.

-- 
Daniel Jacobowitz                           Carnegie Mellon University
MontaVista Software                         Debian GNU/Linux Developer

Re: DWARF-2 and constructors/destructors

[Daniel Berlin]

On Wed, 5 Dec 2001, Daniel Jacobowitz wrote:

> On Wed, Dec 05, 2001 at 10:52:46PM -0500, Daniel Jacobowitz wrote:
> > Suppose you want to call a (non-virtual) method in C++, from something
> > with Dwarf-2 info.  The only way to get the mangled name from the debug
> > info is DW_AT_MIPS_linkage_name.  This isn't, of course, present for
> > constructors/destructors, since the entry in the class definition is
> > for the abstract version.
> >
> > Is it reasonable for the debugger to have to mangle this itself?  The
> > constructor arguments can be arbitrarily complex.  Should there be
> > references in the debug information to the base and complete
> > constructors anywhere?
>
> And I see, in a rather memorable rant, Daniel Berlin saying that
> DW_AT_MIPS_linkage_name is "a hack" (certainly), "not necessary, and
> we're trying to make it go away anyway."

Yes, with which Jason Merrill agreed, IIRC. In fact, he first convinced
me.
:)
I'm just more vitriolic by a small margin.
> Dan, would you mind
> enlightening me on how to cope without it without having a full mangler
> for every supported C++ ABI attached to the Dwarf-2 reader?
Sure.
What exactly do you need mangled names for?
All the methods have real names anyway.
We already demangle them anyway when we lookup symbols, remember.

If you mean for calling constructors, it's abi specific.
Find all the matching constructors, which will be a small number, demangle
them, and pick the right one for your purposes.  The demangler actually
marks what type of constructor/deconstructor it is, the info is just not
propagated to somewhere where gdb sees it (in cp-demangle.c, it's stored
in the demangling_t structure, but we only hand back the string).

>
> As far as I'm concerned, GDB has no business ever mangling a single
> string.
Bingo.
>  Debug info should provide everything that we need.
Yes, it should.

Now go look at check_method_stub and gdb_mangle_name, and try to avoid
hurling.


>
> --
> Daniel Jacobowitz                           Carnegie Mellon University
> MontaVista Software                         Debian GNU/Linux Developer
>

Re: DWARF-2 and constructors/destructors

[Jason Merrill]

>>>>> "Daniel" == Daniel Jacobowitz <dmj+@andrew.cmu.edu> writes:

> Suppose you want to call a (non-virtual) method in C++, from something
> with Dwarf-2 info.  The only way to get the mangled name from the debug
> info is DW_AT_MIPS_linkage_name.

As Dan Berlin points out, The mangled name is already in the symbol table,
if there's actually a definition of the method.  Why does it need to be in
the debug info as well?  To call a non-virtual function, we're only
interested in its address.  The debug information doesn't enter into it.

This applies to all functions, of course, not just ctors/dtors.  The
DW_AT_MIPS_linkage_name stuff was a hack to make dwarf2 work with the
existing stabs-oriented code with minimal effort; it would save a lot of
space if we could get rid of it.

Jason