This will allow to simplify transform_attribute (e.g. no need to handle
references anymore) and to ensure the pre/post/fail calls are made to the
same transformation.
The double conversion `spirit::string -> std::string -> spirit::string` was
happening silently in rules because of `transform_attribute` was returning
`std::string` and implicit constructor in `spirit::string`.
The mistake could be prevented if `spirit::basic_string` constructor from
derived class was explicit, but `utree_cast` uses `is_convertible` while
`is_constructible` is not implemented for C++03 (it is possible, but requires
expression SFINAE support from compiler - N2634).
It was poorly documented and never worked correctly. None of atof, strtof, and
input streams are accepting such values. Fixing it would be a bigger breaking
change than removing it.
Fixes#415, addresses #163 and https://svn.boost.org/trac10/ticket/8699
After #456 `parse_rule` is always called with value of `rule::attribute_type&`
type. By removing attribute type deduction we also will turn any possible bug
in `transform_attribute` from a linkage to compile error.
Fixes#454
This will deal with linking problems when rule attribute type is not the same
as the actual attribute type. It will not harm anyone as the transformation
after parse_rule will just pass-through the attribute because it will be the
same there. It will also do not alter rule_definition parsing so other usages
are still valid.
During overload resolution disallowed signature can be instantiated leading
to firing the static assertation and overload resolution failure. I could
just remove it, but do not want to leave unbounded by value overloads;
replacing to not deducing Attribute with specifying a default parameter will
not work because we want to allow `BOOST_SPIRIT_DECLARE` without prior
`BOOST_SPIRIT_DEFINE` - this requires having default parameters in both
declaration and definition and the standard seems to forbid it (GCC - fine,
MSVC - warning, Clang - error).
Reverts #437, fixes#453.
There is a mistake in `make_attribute` trait that leads to synthesization and
immediately throwing out an attribue, so rule context always was unused_type
when actual attribute is unused_type (added test for this case). Instead of
fixing `make_attribute` trait I just completely removed synthesization in
rule by removing `make_attribute` trait usage (the traits could be removed
after a simple change in 'action' parser). I think the change is better
than fixing the trait because it removes excessive attribute default
construction in every rule and makes rule almost free to use.
When a rule has no attribute (leaves `Attribute` template parameter as default)
the `BOOST_SPIRIT_INSTANTIATE` macro instantiates a `parse_rule` helper function
with a wrong signature because `parse`, `phrase_parse`, and `skip_over` pass as
the attribute `unused` variable which deduces to `unused_type const&` type, but
`BOOST_SPIRIT_INSTANTIATE` instantiates `parse_rule` with `unused_type&` type.
The `partition_attribute` never splits a sequence into zero sized parts.
Removing the specialization simplifies an ongoing patch that will forbid
sequence into plain types parsing.
Not intrusive. Significantly speeds up tests on MSVC (for about 3 times).
Finally solves exceeding build time limits on CI with MSVC compilers.
Did not set up PCH for X3, it is fast as-is (tooks only about 2 minutes).
Repository tests reuse PCH from Qi/Karma tests.
It will allow to test `extract_int` properly.
Replacing `boost::integer_traits` with `std::numeric_limits` will not hurt
performance because even MSVC 9/GCC 4.1/Clang 3.0 folds `min`/`max` calls
to a constant at compile time and since C++11 they are even constexpr.
The `boost/spirit/home/support/detail/lexer/serialise.hpp` header is a private header
and is not used anywhere in Spirit or Boost. No one should be affected by the change.
