Re: [PATCH 1/4] Add named-types-ids to use name ids after the type name instead of numbers.

[Mark Wielaard <mark@klomp.org>]

Hi Matthias,

On Fri, 2020-04-24 at 14:57 +0200, Matthias Maennich wrote:
> On Thu, Apr 23, 2020 at 07:48:12PM +0200, Dodji Seketeli wrote:
> > "Mark J. Wielaard" <mark@klomp.org> a écrit:
> > [...]
> > diff --git a/src/abg-libxml-utils.cc b/src/abg-libxml-utils.cc
> > index 2c46aad8..7370afd5 100644
> > --- a/src/abg-libxml-utils.cc
> > +++ b/src/abg-libxml-utils.cc
> > @@ -249,6 +249,65 @@ escape_xml_string(const std::string& str)
> >   return result;
> > }
> > 
> > +/// Replace the various special characters in a type string as used in
> > +/// a named type-id attribute.
> > +///
> > +/// The characters '<', '>', ''', '"' and ' ' are all replaced by '_'.
> > +/// The characters '&' and '*' at the end of a string are simply dropped,
> > +/// otherwise they are also replaced by an '_'.
> > +///
> > +/// The result is not reversible.
> > +///
> > +//// @param str the input string to read to search for the characters
> > +//// to replace.
> > +////
> > +//// @param replaced the output string where to write the resulting
> > +//// string that contains the replaced characters,
> > +void
> > +escape_xml_string_as_named_type_id(const std::string& str,
> > +				   std::string& replaced)
> > +{
> > +  for (std::string::const_iterator i = str.begin(); i != str.end(); ++i)
> > +    switch (*i)
> > +      {
> > +      case ' ':
> > +      case '<':
> > +      case '>':
> > +      case '"':
> > +      case '\'':
> > +	replaced += '_';
> > +	break;
> > +      case '&':
> > +      case '*':
> 
> Would this not make references types and pointer types
> indistinguishable?

Yes, which is why in the original patch prefix the named-type-id with
the type-kind.

> Consider
> 
>    | template<typename T>
>    | class Test{};
>    |
>    | Test<int&> a;
>    | Test<int*> b;

So "int&" would come out as "ref-int" and "int*" would come out as
"ptr-int".

> While we add underscores until we have a unique match, that will later 
> 
> Also, why would we need to replace those characters at all? Except the
> quote itself, we should be fine?! What am I missing? Can we reuse the
> mangled name? Can replace less of those characters?

As far as I understand '<' and '&' always need to be escaped in XML
attributes. '>' doesn't, but it is really confusing to replace only '<'
and not '>', so to "balance it out" I replaced them both with the same
char. 

> In the textual representation those are replaced differently:
> 
>    name='Test&lt;int&amp;&gt;'
>    name='Test&lt;int*&gt;'

I found that to be really hard to read.
I admit that my tests have all be plain C, for which the replacement
turns out very nicely. I'll try some C++ code to see if I can make the
named-type-ids easier to read. Suggestions welcome. But maybe --named-
type-ids only really makes sense for plain C libraries...?

> > +	if (i + 1 != str.end())
> > +	  replaced += '_';
> > +	break;
> > +      default:
> > +	replaced += *i;
> > +      }
> > +}
> > +
> > +/// Replace the various special characters in a type string as
> > used in
> > +/// a type-id attribute.
> > +///
> > +/// The characters '<', '>', ''', '"' are all replaced by '_'.
> > +/// The character '&' is replaced by the string "-ref".
> 
> Is this really what is happening? I seem to miss the part where & is
> transformed into '-ref'.

Bad comment, the same function right above has the correct comment.
Instead of prefixes I originally used postfix -ref and -ptr, but that
didn't distinguish other types, which don't have special characters in
their names (like typedefs, which are often named after the underlying
type). I'll fix up the comment.

> > diff --git a/src/abg-writer.cc b/src/abg-writer.cc
> > index 74fa1a8c..5f85810b 100644
> > --- a/src/abg-writer.cc
> > +++ b/src/abg-writer.cc
> > @@ -127,6 +127,29 @@ public:
> >     ABG_ASSERT(env);
> >     return env->intern(o.str());
> >   }
> > +
> > +  /// Return a unique string representing a type representation.  The
> > +  /// returned string will be made unique by postfixing underscores if
> > +  /// necessary.
> > +  ///
> > +  /// @param type to create an unique id string for
> > +  interned_string
> > +  get_named_type_id(const type_base* type) const
> > +  {
> > +    string named_type_id =
> > +      xml::escape_xml_string_as_named_type_id(get_pretty_representation(type,
> > +									/*internal=*/true));
> > +
> > +    /* We want to make sure the id is unique. See if it is already
> > +       interned in this environment, if it is, it isn't unique and we
> > +       add some underscores to it till it is.  */
> > +    const environment* env = get_environment();
> > +    ABG_ASSERT(env);
> > +    while (env->is_interned_string(named_type_id))
> > +      named_type_id = named_type_id + "_";
> 
> Postfixing the underscore has the advantage that this is not
> reproducibly producing the same output for the same input. That would
> defeat (a bit) the purpose of that patch to stabilize the output.

I am afraid I don't fully understand what you are suggesting here.

There is an issue with anonymous types. Since they have the same names.
I don't have a good solution for that, but think it might be possible
to disambiguate them in most cases.

Another issue is using the exact same type name defined in different
translation units that are not the same type. That issue violates ODR
for C++, but can happen for C code. If it happens, it does depend on
the order we see the translation unit. But in the cases I am interested
in you would use --named-type-ids together with --drop-private-types.
Which makes it more likely that such internal same name, but different
type, issues pop up.

Cheers,

Mark