gdb 10.0 fails to examine any global variables in D programs

[Zixing Liu]

[ Impact ]

 * GDB 10.0 introduced a regression where it cannot inspect any global variables in any D programs compiled by any D compiler.
 * LDC2 and GDC upstream stated Focal does not have such a problem and stuck to this release for their test images.

[ Test Plan ]

Considering the following D program:

```
module t;

class uv {
    int i;
}

__gshared uv i;
int main() {
    i = new uv();
    return 0; // #break
}
```

If you build it using `gdc -g -O0 t.d -o t` or `ldc2 -o t.o t.d -g`, run the GDB using the following commands ...

```
b t.d:10
p t.i
```

... you will notice GDB will complain that "'t.i' has unknown type; cast it to its declared type."

[ Where problems could occur ]

  * The fix consists of a single line change to the demangler. The worst-case scenario would be breaking the demangling functionality of other programming languages. However, the newer D ABI uses a symbol mangling scheme that is very difficult to confuse with other programming languages.
  * Incorrect symbol de-mangling may also cause user confusion. However, the patch fixed a fundamental usability issue.

[ Other Info ]

  * Initial discussion in the LDC2 bug tracker: https://github.com/ldc-developers/ldc/issues/4389

[Zixing Liu (liushuyu-011)]

Created attachment 15447
Minimal reproducible example (D source code)

Hi there,

I have discovered a D class type resolution regression since GDB 10 (it was working properly in GDB 9).

Please find the reproducer in the file attached. You can compile the source code using either GDC or LDC2 (both can reproduce this issue):

gdc -g -O0 t.d -o t

To reproduce the bug, set the breakpoint to t.d:10 and examine the `i` global variable (using `p t.i`). GDB will then complain that "'t.i' has unknown type; cast it to its declared type."

I found a workaround to this problem by using `p (uv*)_D1t1iCQf2uv`, but obviously, this is non-ideal: the developer may not know the mangled variable name easily.

Since this looks like a regression, I did a simple `git bisect` that led me to this commit: https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=727b7b1864973c2645a554727afd0eaf1303673a. I am unsure how demangle changes could affect the type resolution (maybe it's because GDB got confused about which entity to decode?)

[Tromey-b (tromey-b)]

Computing the "physname" decides that the symbol should be named:

(top-gdb) p physname
$37 = 0x2c70490 "_D1t1iCQf2uv"

... which is plainly wrong.

[Tromey-b (tromey-b)]

Ok, so while gdb's symbol reader is still really wrong here --
this physname stuff is pretty broken -- the root cause of this
particular bug is also that libiberty does not auto-demangle
D symbols.

You can see this on the command line:

prentzel. c++filt _D1t1iCQf2uv
_D1t1iCQf2uv
prentzel. c++filt -s dlang _D1t1iCQf2uv
t.i

To my eye this seems to be an oversight in cplus-dem.c,
where the D code follows some earlier code; but perhaps
the author didn't realize that the reason the Ada
code does not check AUTO_DEMANGLING is that the
Ada encoding isn't unambiguous in the way the D encoding is.

[Tromey-b (tromey-b)]

Testing a patch.

[Tromey-b (tromey-b)]

https://sourceware.org/pipermail/gdb-patches/2024-March/207686.html

[Cvs-commit (cvs-commit)]

The master branch has been updated by Tom Tromey <email address hidden>:

https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=b1741ab0dafd899889faab6e862094a325a6b83c

commit b1741ab0dafd899889faab6e862094a325a6b83c
Author: Tom Tromey <email address hidden>
Date: Sat Mar 30 13:48:30 2024 -0600

    libiberty: Invoke D demangler when --format=auto

    Investigating GDB PR d/31580 showed that the libiberty demangler
    doesn't automatically demangle D mangled names. However, I think it
    should -- like C++ and Rust (new-style), D mangled names are readily
    distinguished by the leading "_D", and so the likelihood of confusion
    is low. The other non-"auto" cases in this code are Ada (where the
    encoded form could more easily be confused by ordinary programs) and
    Java (which is long gone, but which also shared the C++ mangling and
    thus was just an output style preference).

    This patch also fixed another GDB bug, though of course that part
    won't apply to the GCC repository.

    Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=31580
    Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=30276

    libiberty
            * cplus-dem.c (cplus_demangle): Try the D demangler with
            "auto" format.
            * testsuite/d-demangle-expected: Add --format=auto test.

[Tromey-b (tromey-b)]

Fixed.