Re: [PATCH][gdb/tui,c++17] Fix NULL string_view in tui_partial_win_by_name

[Tom de Vries <tdevries@suse.de>]

On 7/6/20 4:47 AM, Simon Marchi wrote:
> On 2020-07-05 5:40 p.m., Tom de Vries wrote:
>> Hi,
>>
>> When building gdb with CFLAGS=-std=gnu17 and CXXFLAGS=-std=gnu++17 and running
>> test-case gdb.tui/new-layout.exp, we run into:
>> ...
>> UNRESOLVED: gdb.tui/new-layout.exp: left window box after shrink (ll corner)
>> FAIL: gdb.tui/new-layout.exp: right window box after shrink (ll corner)
>> ...
>>
>> In a minimal form, we run into an abort when issuing a winheight command:
>> ...
>> $ gdb -tui -ex "winheight src - 5"
>>    <tui stuff>
>> Aborted (core dumped)
>> $
>> ...
>> with this backtrace at the abort:
>> ...
>> \#0  0x0000000000438db0 in std::char_traits<char>::length (__s=0x0)
>>      at /usr/include/c++/9/bits/char_traits.h:335
>> \#1  0x000000000043b72e in std::basic_string_view<char, \
>>    std::char_traits<char> >::basic_string_view (this=0x7fffffffd4f0, \
>>    __str=0x0) at /usr/include/c++/9/string_view:124
>> \#2  0x000000000094971b in tui_partial_win_by_name (name="src")
>>      at src/gdb/tui/tui-win.c:663
>> ...
>> due to a NULL comparison which constructs a string_view object from NULL:
>> ...
>>    657  /* Answer the window represented by name.  */
>>    658  static struct tui_win_info *
>>    659  tui_partial_win_by_name (gdb::string_view name)
>>    660  {
>>    661    struct tui_win_info *best = nullptr;
>>    662
>>    663    if (name != NULL)
>> ...
>>
>> In gdbsupport/gdb_string_view.h, we either use:
>> - gdb's copy of libstdc++-v3/include/experimental/string_view, or
>> - the standard implementation of string_view, when built with C++17 or later
>>   (which in gcc's case comes from libstdc++-v3/include/std/string_view)
>>
>> In the first case, there's support for constructing a string_view from a NULL
>> pointer:
>> ...
>>       /*constexpr*/ basic_string_view(const _CharT* __str)
>>       : _M_len{__str == nullptr ? 0 : traits_type::length(__str)},
>>         _M_str{__str}
>>       { }
>> ...
>> but in the second case, there's not:
>> ...
>>       __attribute__((__nonnull__)) constexpr
>>       basic_string_view(const _CharT* __str) noexcept
>>       : _M_len{traits_type::length(__str)},
>>         _M_str{__str}
>>       { }
>> ...
>>
>> Fix this by removing the NULL comparison altogether.
>>
>> Build on x86_64-linux with CFLAGS=-std=gnu17 and CXXFLAGS=-std=gnu++17, and
>> tested.
>>
>> Any comments?
> 
> This is the second string_view problem in a short amount of time, the other being:
> 
>   https://sourceware.org/bugzilla/show_bug.cgi?id=26183
> 
> When importing the string_view support, I ended up copying experimental/string_view
> (which is a version from before it was a standard) from libstdc++, rather than
> std/string_view.  In this message:
> 
>   https://sourceware.org/legacy-ml/gdb-patches/2018-04/msg00111.html
> 
> I said it was because "it is easier to adapt to the c++ standard we follow", but I
> don't remember the specific reason.

Sounds like a good rationale to me.

> I think we should change our string_view to
> be an adapted copy of std/string_view instead, to minimize the chances of differences
> with the standard version.  I started to do it the other day (but did not finish
> adapting the tests), and didn't see any blocker.
> 

I suppose it's a trade-off, but I agree ideally we use std/string_view.

> I still think your patch is good (it would be necessary in any case).

Ack, committed, and thanks for the explanation.

- Tom