Expanded on the requirements for resource types and values supported by
unallocated_resource.
Added a note in the INVALID_HANDLE_VALUE example that says it relies on
MSVC extension and other compilers may not accept it.
The support required to be permissive wrt. is_allocated being callable
with a resource argument. This could lead to unexpected behavior, where
the user provided a seemingly conforming resource traits but
unique_resource would still reject them for some reason (e.g. if
is_allocated is not callable with the const resource value, as tested
by unique_resource) and work in "reduced" resource traits mode. This
would change the behavior of the APIs that accept resource values as
an argument (e.g. constructing unique_resource with a resource value
would always produce an object in the allocated state, while the user
would expect it to depend on the resource value).
Removing the support for "reduced" resource traits makes the behavior
more predictable. If the user specified resource traits, it either
compiles and works according to the resource traits, or fails to compile
if the resource traits are somehow not conforming.
The support for customizing default-constructed resource value is not
important enough. It can be worked around by either implementing proper
resource traits, or by specifying the default value in
make_unique_resource_checked.
Additionally, modified unallocated_resource to relax the argument type
of is_allocated and the types of the unallocated values. This can be
useful when the types of the unallocated resource values don't match
the actual resource type (e.g. std::nullptr_t vs. typed pointers).
This avoids potentially assigning the default to the resource value, which
may be expensive and is pointless, as the resource is about to be destroyed
anyway.
Removed Boost.ScopeExit from the table on the front page. Comparison with
Boost.ScopeExit is in a separate section, so this table duplicated it
to some degree and detracted focus from introducing Boost.Scope.
Added an example with unique_resource to the front page.
Other wording and example improvements.
Instead of silently switching to the unique_resource implementation
based on the "allocated" flag if the resource traits don't conform
to the requirements, issue hard compilation errors. This should better
protect against mistakes, where the user assumes unique_resource
behaves as if it is using the resource traits while it actually doesn't.
Improved documentation of requirements for resource traits and documented
the "reduced" form of resource traits, where the traits only provide
the default resource value but otherwise don't change unique_resource
behavior.
One person during review complained that the library implementation
used typedefs instead of C++11 using type aliases. Not an essential
difference, but if it helps code readability, let it be.
The new row of examples illustrate explicit deactivation of the scope
guards, as opposed to automatic failure detection via scope_fail.
Also improved the other examples to look more realistic.
Closes https://github.com/Lastique/scope/issues/13.
If the resource type is nothrow move-constructible and the deleter type
is not, and the deleter's copy constructor throws during unique_resource
move constructor, then move the move-constructed resource back to the
original object instead of invoking the deleter on it. This leaves the
move source in its original state rather than in unallocated state,
which means unique_resource move constructor now provides strong
exception guarantee.
