Re: [PATCH] gdb/solib: pass lm_info, original_name and name to solib constructor

[Simon Marchi <simark@simark.ca>]

On 2025-06-03 19:07, Guinevere Larsen wrote:
>> +    sos.emplace_back (std::move (info), info->name, info->name);
> This is a bad idea. I double checked with c++ folk and while no
> compilers warn about this, if the compiler evaluates things right to
> left (like gcc) this works, but if they evaluate left to right (like
> clang), then info->name will be de-referencing null and segfault. This
> might also depend on ABI, but the end result is the same: fragile at
> best. we'll need something that saves or moves info->name to pass it
> to the emplace_back constructor

Wow, good catch, thanks.  Actually, with C++17, I think the order of
evaluation has been specified from left to right.  So, I think it should
crash regardless of the compiler.  I don't think that the testsuite
exercises that code though on Linux.

A more long-term fix would be to move the name out of lm_info_target (it
doesn't belong there anyway).  But for now I'll change it like so in my
patch:


  for (lm_info_target_up &info : library_list)
    {
      /* Move NAME to a local variable to avoid reading INFO->NAME after having
	 moved info.  */
      std::string name = std::move (info->name);
      sos.emplace_back (std::move (info), name, name);
    }

It seems to be the only occurrence of this problem in the patch.

Simon