Remove some variables in favor of using gdb::optional

  On 2019-08-21 8:44 p.m., Simon Marchi wrote:
> On 2019-08-21 3:38 p.m., Pedro Alves wrote:
>> std::optional<bool> is evil.  :-)
>>
>> E.g. it's very easy to write (or miss converting)
>>
>>  if (is_static)
>>
>> and not realize that that is doing the wrong thing.
>>
>> https://www.boost.org/doc/libs/1_57_0/libs/optional/doc/html/boost_optional/tutorial/a_note_about_optional_bool_.html
>>
>> Not saying to change the code (*), but seeing it gives me the creeps, so I couldn't resist.  :-)
>>
>> * - a yes/no/unknown tristate type a-la boost::tribool would be nice
>>     to have, IMO.  It could be used more clearly in these situations
>>     and could supersede auto_boolean.
>>
>> Thanks,
>> Pedro Alves
> 
> Oh, thanks for pointing out.  I had never realized this but it makes sense.  There should be a compiler
> warning about it!  Or maybe it would belong to a linter.
> 
> I'll look into adding a tristate bool and fixing it.
> 
> Simon
> 

I started looking into it, seeing if I could implement something like boost's tribool,
but then wondered, why don't we use boost directly?

Its license is compatible with the GPL [1], which from what I understand means that it
can be used to build a GPL program.

[1] https://www.gnu.org/licenses/license-list.html#boost

Boost is very widely available, so I'd prefer if we could depend on the system-installed
boost, or have the user provide it, but I would understand if others didn't want to add
that burden to those who build GDB.  So an alternative would be to carry a version of boost
(maybe just the files that we need) in our repo.  This would have the advantage that everybody
would build with the same version, hence less chances of build failures with particular
combinations of compilers and boost versions, or using features from a too recent boost.

What do you think?

Here's what a patch to remove the gdb::optional<bool> would look like if we used boost:tribool:


From 440b941b9b839943ccabf90c6370450580e92700 Mon Sep 17 00:00:00 2001
From: Simon Marchi <simon.marchi@efficios.com>
Date: Thu, 22 Aug 2019 19:13:13 -0400
Subject: [PATCH] gdb: Use boost:tribool for is_static in
 dw2_debug_names_iterator::next

---
 gdb/dwarf2read.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
  

On 8/23/19 12:36 AM, Simon Marchi wrote:
> On 2019-08-21 8:44 p.m., Simon Marchi wrote:
>> On 2019-08-21 3:38 p.m., Pedro Alves wrote:
>>> std::optional<bool> is evil.  :-)
>>>
>>> E.g. it's very easy to write (or miss converting)
>>>
>>>  if (is_static)
>>>
>>> and not realize that that is doing the wrong thing.
>>>
>>> https://www.boost.org/doc/libs/1_57_0/libs/optional/doc/html/boost_optional/tutorial/a_note_about_optional_bool_.html
>>>
>>> Not saying to change the code (*), but seeing it gives me the creeps, so I couldn't resist.  :-)
>>>
>>> * - a yes/no/unknown tristate type a-la boost::tribool would be nice
>>>     to have, IMO.  It could be used more clearly in these situations
>>>     and could supersede auto_boolean.
>>>
>>> Thanks,
>>> Pedro Alves
>>
>> Oh, thanks for pointing out.  I had never realized this but it makes sense.  There should be a compiler
>> warning about it!  Or maybe it would belong to a linter.
>>
>> I'll look into adding a tristate bool and fixing it.
>>
>> Simon
>>
> 
> I started looking into it, seeing if I could implement something like boost's tribool,
> but then wondered, why don't we use boost directly?
> 
> Its license is compatible with the GPL [1], which from what I understand means that it
> can be used to build a GPL program.
> 
> [1] https://www.gnu.org/licenses/license-list.html#boost
> 
> Boost is very widely available, so I'd prefer if we could depend on the system-installed
> boost, or have the user provide it, but I would understand if others didn't want to add
> that burden to those who build GDB.  

Yeah, boost would be a too-big dependency, IMO.  It's huge.

> So an alternative would be to carry a version of boost
> (maybe just the files that we need) in our repo.  This would have the advantage that everybody
> would build with the same version, hence less chances of build failures with particular
> combinations of compilers and boost versions, or using features from a too recent boost.
> 
> What do you think?

I would really not like to carry boost.  It's very big and intertwined.  I think the
costs outweighs the benefits here, because we'd be pulling in a huge codebase when we
only need minimal things.  I'd also like to move more in the direction of sharing more
code across the toolchain (gdb, gcc, gold, etc.) and depending on boost would
certainly prevent that.  LLVM doesn't depend on boost either, for example.  Given
that it's a C++ project from the get go, I think that's telling.  Most certainly for
the same reasons.

On reusing parts, when I was working on the original gdb::unique_ptr emulation for
C++98, before we upgraded to C++11, I looked at reusing boost's version.  What I
found was a huge horrible mess, with lots of incomprehensible #ifdefery to support
compilers that no one cares about anymore (old Borland C++, etc.), impossible to
decouple.  Another issue is that a boost API may have been originally designed
with C++98 in mind, and it'd be done differently with C++11 and later in mind.
It might have been the case here, for example note how boost's tribool uses
the C++98 safe bool idiom, instead of C++11 operator bool().

> 
> Here's what a patch to remove the gdb::optional<bool> would look like if we used boost:tribool:


> 
> 
> From 440b941b9b839943ccabf90c6370450580e92700 Mon Sep 17 00:00:00 2001
> From: Simon Marchi <simon.marchi@efficios.com>
> Date: Thu, 22 Aug 2019 19:13:13 -0400
> Subject: [PATCH] gdb: Use boost:tribool for is_static in
>  dw2_debug_names_iterator::next
> 
> ---
>  gdb/dwarf2read.c | 7 ++++---
>  1 file changed, 4 insertions(+), 3 deletions(-)
> 
> diff --git a/gdb/dwarf2read.c b/gdb/dwarf2read.c
> index de9755f6ce30..915e0b25cea1 100644
> --- a/gdb/dwarf2read.c
> +++ b/gdb/dwarf2read.c
> @@ -90,6 +90,7 @@
>  #include <forward_list>
>  #include "rust-lang.h"
>  #include "gdbsupport/pathstuff.h"
> +#include <boost/logic/tribool.hpp>
> 
>  /* When == 1, print basic high level tracing messages.
>     When > 1, be more verbose.
> @@ -5843,7 +5844,7 @@ dw2_debug_names_iterator::next ()
>        return NULL;
>      }
>    const mapped_debug_names::index_val &indexval = indexval_it->second;
> -  gdb::optional<bool> is_static;
> +  boost::tribool is_static(boost::indeterminate);

Here I imagine we'd more naturally want to treat a tribool like
a built-in, similar to an enum class that accepts {true, false, unknown},
so we'd write:

  tribool is_static = true;
  tribool is_static = false;
  tribool is_static = tribool::unknown;

("indeterminate" is really a mouthful, hence "unknown".)

with tribool::unknown being e.g., a constexpr global.

>    dwarf2_per_cu_data *per_cu = NULL;
>    for (const mapped_debug_names::index_val::attr &attr : indexval.attr_vec)
>      {
> @@ -5910,10 +5911,10 @@ dw2_debug_names_iterator::next ()
>      goto again;
> 
>    /* Check static vs global.  */
> -  if (is_static.has_value () && m_block_index.has_value ())
> +  if (!boost::indeterminate(is_static) && m_block_index.has_value ())

and here:

+  if (is_static != tribool::unknown && m_block_index.has_value ())

Here's a quick-prototype starter implementation.  It's missing sprinkling
with constexpr, inline, noexcept, relational operators (op|| and op&& ?),
unit tests, etc., but should show the idea.

#include <stdio.h>

class tribool 
{
private:
  struct unknown_t {};
public:
  /* Keep it trivial.  */
  tribool () = default;

  tribool (bool val)
    : m_value (val)
  {}

  operator bool () const
  {
     return m_value == 1;
  }

  bool operator== (const tribool &other)
  {
    return m_value == other.m_value;
  }

  tribool operator! ()
  {
    if (m_value == 0)
      return true;
    else if (m_value == 1)
      return false;
    else
      return unknown;
  }

  static tribool unknown;

private:
  tribool (unknown_t dummy)
    : m_value (-1)
  {}

  int m_value; // -1 for unknown.
};

tribool tribool::unknown (unknown_t{});

int
main ()
{
  tribool b1 = true;
  tribool b2 = false;
  tribool b3 = tribool::unknown;
  tribool b = b3;

  if (b)
    printf ("then\n");
  else if (!b)
    printf ("else\n");
  else 
    printf ("unknown\n");
  
  return b1;
}

Would you like to run with this?

Thanks,
Pedro Alves