Re: Regression handling linkage vs source name

[Doug Evans <dje at google dot com>]

On Tue, Feb 21, 2012 at 4:41 PM, Michael Eager <eager@eagercon.com> wrote:
> There were changes to dwarf2_physname() around July 1, 2011, which
> cause it to return a different value. ?Arguably, the changes make
> the function work better, returning the linkage name for a symbol where
> it previously returned the source name.

Tangential data point: Outside of dwarf2read.c, gdb generally uses
"physname" to mean linkage name.

I'm not sure dwarf2_physname ever returned the linkage name (for c/c++).
I know there's been some changes in how it computes the source name.

> But since the source name
> is overwritten by the linkage name in the symbol entry, gdb's
> behavior is different, and, IMO, incorrect.

Can you elaborate on what you mean by overwritten?
[It's not what I see.]

> Here is the test case:
>
> $ cat t.c
> extern int foo(void) asm ("xfoo");
>
> int foo(void) {}
>
> int main (void)
> {
> ?foo ();
> }
>
> $ gdb-before a.out
> ...
> (gdb) b foo
> Breakpoint 1 at ...
>
> $ gdb-after a.out
> ...
> (gdb) b foo
> Make breakpoint pending on future shared library load? (y or [n])? n
> (gdb) b xfoo
> Breakpoint 1 at ...
>
>
> The symbol "foo" is no longer present in gdb symbol table
> so trying to use the name specified in the source fails.
> Before the change, breakpoint, backtrace, and other commands
> display the symbol name as in the source (and in the DWARF DIE).
> After the change, the linkage name is displayed instead of the
> source name.

Even pre-dwarf2_physname gdb's have this problem.
Maybe you tried a 7.x variant that I didn't (I think I tried 6.8, 7.0,
7.1, and 7.4 - at the time I didn't have easy access to 7.2,7.3 and
didn't want to go to trouble of building them).

The problem (or at least one of the problems), as I see it, is that
the API for creating symbols is broken.
gdb does (or at least can) record both the linkage and source names
(it does this for "minsyms", e.g. ELF symbols).  But the way to set a
symbol's name is via symbol_set_names and it only takes a linkage name
(though the dwarf2read.c further compounds the problem by passing it
the source name - or more accurately the "search" name).
symbol_set_names then tries to demangle the name and will record that
name as well if the demangling succeeds.

> Before the change, dwarf2_physname() calls dwarf2_compute_name()
> which returns the symbol name if the language is not C++, Java,
> or Fortran, even if the DIE has a DW_AT_linkage_name attribute.
> After the change, dwarf2_physname() returns the DW_AT_linkage_name.
>
> Since gdb doesn't keep both the source name and the linkage
> name in the symbol table (a design flaw, IMO) what is the
> right way to get the previous behavior, where gdb recognizes
> the symbol names from the source file?

We need to have dwarf2read.c store both names (linkage name and dwarf
name). [More changes may be required beyond that, but I think that's a
start.]  Your test program shows that we can't assume we can simply
demangle the linkage name to get the source name.