Remove some variables in favor of using gdb::optional

From: Simon Marchi <simon.marchi@efficios.com>

On 2019-08-23 11:35 a.m., Pedro Alves wrote:
>> 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.

True, libbost1.67-dev on Debian unpacks to 150 MB of header files.  It would be nice to be able to
select which features one actually need and check-in just the necessary files (a bit like gnulib).
I bet that for tribool, it would be just a few files.

> 
>> 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().

Ok, all good points.

>> 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;

Yes that sounds good.

> 
> ("indeterminate" is really a mouthful, hence "unknown".)
I also wasn't sure about this, what the third state should be, more on that below.

> 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?

So I wasn't sure about what the third state should be.  I think it depends on
the particular context.  In different contexts, it could mean "unknown", "unspecified",
"auto", "don't care", etc.  There's no one-size word that fits all case, so I don't really
like the idea of having just one word and have it represent poorly what we actually mean.

That lead me to think, if we want to represent three states and if the states are
specific to each use case, why not just define an enum and be explicit about it?

A bit like why I prefer defining an explicit type with two fields rather than using
std::pair: the "first" and "second" members are not very descriptive.

Here's a patch that does that.  What do you think?

From ba180d647f6ceb69b9a01c46807c102439b8cae1 Mon Sep 17 00:00:00 2001
From: Simon Marchi <simon.marchi@efficios.com>
Date: Fri, 23 Aug 2019 14:53:59 -0400
Subject: [PATCH] dwarf2read: replace gdb::optional<bool> with enum

gdb::optional<bool> is dangerous, because it's easy to do:

  if (opt_bool)

when you actually meant

  if (*opt_bool)

or vice-versa.  The first checks if the optional is set, the second
checks if the wrapped bool is true.

Replace it with an enum that explicitly defines the three possible
states.

gdb/ChangeLog:

	* dwarf2read.c (dw2_debug_names_iterator::next): Use enum to
	represent whether the symbol is static, dynamic, or we don't
	know.
---
 gdb/dwarf2read.c | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

Remove some variables in favor of using gdb::optional

Commit Message

Comments

Patch