[RFA] dwarf2_physname

Keith Seitz keiths@redhat.com
Tue Sep 15 15:53:00 GMT 2009


On 09/15/2009 06:32 AM, Daniel Jacobowitz wrote:

> Oh, I see.  Is there any other differentiation in the ICC output
> between the two relevant cases?  Namely bar and baz in:
>
> class foo {
>    void bar();
>    static void baz(foo*);
> };

In this case, the DWARF output looks pretty much as expected. There is 
only one DIE defining foo::baz, and that is in the class definition. One 
quirk: the parameter's DW_AT_name is "$01".

I played with this a bit, and it seems that only ctors and dtors have 
DIEs outside the class definition (ignoring "normal" C functions). 
However, in every case I tried, ICC does not output DW_AT_artificial for 
ANY class method's first parameter. Does ICC not use hidden parameter 
passing to implement method calls? It seems odd that it would do this 
for ctors/dtors and no other methods.

In any case, once again, DW_AT_name for method formal parameters is also 
generated ("$02", "$03", etc). It seems that the only time I see a 
useful DW_AT_name for anything function-like is for the ctor/dtor (all 
parameters, not just hidden "this") and C functions.

[Anecdote: Every time I run readelf -w on an ICC-generated object, it 
complains of "Bogus end-of-siblings marker" -- perhaps there are some 
bytes in the debuginfo (which readelf/gdb cannot read/interpret) which 
helps mitigate some of these problems?]

>  From discussions on the dwarf-discuss list, I've gathered that pretty
> much every debugger is full of compiler-specific quirking.  In
> practice, to provide a good user experience, we have to interpret the
> DWARF in our best understanding of each compiler's intentions rather
> than just through the standard.

Sure, that doesn't really surprise me. But these three quirks 
(conflicting DIEs for same object, missing DW_AT_artificial, nonsensical 
DW_AT_name) would seem truly unnecessary and fairly trivial to fix.

> Of course, that relies on someone wanting to do the work for ICC support.

It may just be useful to provide (yet another) switch to turn on 
DW_AT_MIPS_linkage_name processing (and disable the proposed physname 
work). We could do this switch manually (set dwarf-linkage-name on?) 
and/or by producer (turn it on for ICC by default).

Keith



More information about the Gdb-patches mailing list