Re: [RFA] dwarf2_physname

[Michael Eager <eager@eagercon.com>]

Daniel Jacobowitz wrote:
> On Tue, Sep 01, 2009 at 04:17:56PM -0700, Michael Eager wrote:
>> I think that this is a problem with how the template instantiation is
>> described.
> 
> For your example DWARF:
> 
>> I think that there should be something like
>>    DW_TAG_template_value_parameter
>>      DW_AT_name : Foo
>>      DW_AT_type : <int *>
>> as a child of  of f<Foo>.
> 
> I believe DW_AT_name would be S here and not Foo - the name by which
> the argument is known inside the instantiation.  It's representing Foo
> that's a problem; it has to go in the DW_AT_const_value attribute.

Yeah.  I was trying to figure out whether it was the formal or
actual parameter.  I wanted "Foo" to appear somewhere.  I'm not
sure DW_AT_const_value is really appropriate either.

> Here's a trickier example:
> 
> template<const char *S> int f()
> {
>   return S[0];
> }
> 
> template <int I> struct S {};
> 
> template<int I, const char *STR> int g(S<I + 1 + sizeof(STR)>*)
> {
>   return 0;
> }
> 
> char Foo[3];
> char Bar[3];
> 
> int main()
> {
>   return f<Foo>() + f<Bar>() + g<5, Foo>(0);
> }
> 
> 0000000000000000 W _Z1gILi5EXadL_Z3FooEEEiP1SIXplplT_Li1EszT0_EE
> 0000000000000000 W int g<5, &Foo>(S<((5)+(1))+(sizeof (&Foo))>*)
> 
> A patch that can't handle this may still be an improvement - but I'd
> like to know how we're approaching this problem, and whether (A) the
> lack of compilers generating adequate data, and (B) to the best of my
> knowledge, the lack of DWARF constructs for such expressions, is
> going to interfere with debugging of heavily templated programs.

The DWARF Committee had some discussions about how templates
are described in DWARF.  Only the instantiations are described, not
the declaration.  DWARF tries (and sometimes succeeds) in providing
enough info for the instantiation that a debugger can infer either
the template declaration or how it was instantiated.  This breaks
down in several areas.

There was brief discussion about replacing the current description of
templates.  The existing scheme was developed in the early 90's with
little real-world experience.  C++ has evolved significantly since
then.  On the other hand, I'd be concerned that we would come up with
a new description which requires a significant effort to implement and
that neither compiler writers nor debugger developers want to adopt it.
Sort of like the resistance to the transition from DWARF 1 to DWARF 2.

In any case, changes in template descriptions will not be considered
for DWARF 4, but possibly for DWARF 5.  Suggestions always welcome.


-- 
Michael Eager	 eager@eagercon.com
1960 Park Blvd., Palo Alto, CA 94306  650-325-8077