GitLab

This changes the `right` and `tolerance` argument into more complex
objects that allow both a literal, or special objects returned by
`search_right()`, `absolute_tolerance(.001)`, or
`relative_tolerance(.1)`.

This allows the much more descriptive syntax:

    constrained_maximum_search(f, 0, search_right(), relative_tolerance(.001))

This also reintroduces the ability to specify an absolute tolerance
level, which was there once (as a separate argument) but got lost along
the way.

This commit also changes the `domain_t` argument to be no longer
deduced.  Previously it was deduced from `left`, which would have made
the above example rather broken: `domain_t` would have been `int` (from
the 0) rather than `double` (which would require 0.0).  Overriding
`domain_t` now requires an explicitly template parameter:

    constrained_maximum_search<float>(f, 0, search_right(), relative_tolerance(.001))

This allows readLock/writeLock to take any number of arguments, where
each argument can be a container or a member.  Previously you could
pass a single container, or any number of members, but not combinations
of both or multiple containers.

This outputs something like `Member[n]` or `Firm[n]`, and can be
overridden on a per-subclass basis.  It also works for shared members
and weak members.

No code changes: just removing extra class body indentation and comment
rewrapping.  Applies to:

Member.hpp
Good.hpp
Market.hpp
Firm.hpp

There are still more left, but these in particular are changed in the
subsequent commit.

Previously `eris::range(...)` only accepted a `std::pair` argument.
This commit adds support for also accepting a code-wise pair (e.g.
`eris::range(it1, it2)`).

- Specifying NaN for automatic `right` selection resulted in an infinite
  loop when `left` was 0.  Changed to start searching at 1.0 when that
  occurs.
- The `constrained_minimum_search` didn't implement automatic `right`
  selection; fixed.
- Minor tweaks to calculation to avoid promotion to double; the main
  change was introducing a `half()` function to handle division by 2
  without resorting to actual division for floating points while still
  working for non-floating point domain types.
- Fixed a typo in the static_assert for a `bool` return type in
  `constrained_maximum_search`.

`using namespace eris;` from within a namespace doesn't resolve the
ambiguity for `time_t`/`id_t` -- they need an explicit symbol import, or
full qualification.

The previous change to single_peak_search which deduced the domain type
from the arguments was pretty undesirable: if you specified both left
and right as integers, you'd get a search over the integer domain.

This removes the type deduction, instead using an optional template
parameter which defaults to double and isn't deduced, thus allowing
custom types to be used, but falling back to double if not specified.

This also fixes the test code to remove the tests that specified both
relative and absolute tolerances (the latter was removed).

Using a `std::set` to sort is typically less efficient than building a
vector then sorting it (once).

This lets external code use `eris::id_t` and `eris::time_t` which is
far nicer than `eris::eris_id_t`.

The existing `eris_id_t` and `eris_time_t` are kept as deprecated
aliases for the new ones.

The operators weren't returning `*this`; this somehow magically worked
properly under gcc; clang noticed and generated an illegal instruction.

WeakMember<T> is a wrapper around `std::weak_ptr<T>`, with
SharedMember<T> integration.

This adds support for automatically finding `right` based on `left`: if
`right` is given as NaN, a search starts at `2*left` (if `left`
positive) or `-left` (if `left` negative) and continuously doubles the
potential `right` value until a constraint-violating value is found;
this then forms the RHS limit.

- std::pair was missing a load_from/store_to implementation
- reimplemented std::pair support as std::tuple support, with std::pair
  being just a specialization of the tuple serializer.
- De-protected load_from/store_to; they require friend classes all over
  the place because they aren't really "protected".
- remove serializer.cpp in favour of just inlining its three small
  functions in the .hpp.

This lets you numerically find the largest or smallest value satisfying
some condition.

For example, to determine the highest price that has at least `x` sales:

    auto quantity_demanded = [](double p) { return (int)(99 - 2*p); };
    int x = 10;
    double min = 0.0, max = 100.0;
    auto r = eris::constrained_maximum_search(
        [&](double p) { return quantity_demanded(p) >= x; },
        min, max, 0 /* maximum precision */);
    std::cout << std::setprecision(16) << "p = " << r.arg << "\n";
    // p = 44.5

With less precision for `tol_rel` (e.g. the default, 1e-10) you'll
generally get some number slightly smaller than 44.5.

This makes single_peak_search templated (so that it doesn't have to be
implemented with double).  It also updates the return type struct with
templated types, a constructor, and an iterator counter.

The first method MemberID arguments already accept shared members
arguments.

This significantly simplifies the member locking implementation since we
no longer have to worry about tracking the number of active locks;
instead read locks are shared locks and write locks are exclusive locks;
the stl implementation can worry about the details beyond that.

This also disallows recursive member locks; instead calling code has to
manage itself to avoid double-locking a member.  Note also that
switching a write lock to a read lock is no longer guaranteed to be
non-blocking as there is no way to downgrade an exclusive lock to a
shared lock.  boost::upgrade_mutex could allow this, but this importance
of this is minor and it seems nicer to get rid of the boost dependency.

It also changes the main Simulation run lock to be the same stl
implementation; previously it was using the boost implementation.

With that, the boost thread, date_time, and atomic library components
are no longer required.

Previously it was hard-coded to use `build`.

Add a second parameter (priority) to the callback constructors and
provide an override for the associated priority.

This also simplifies the structure by moving the constructors and
function/priority storage into a common base class.

Everything also gets inlined as every method here is extremely simple.

- and/or -> &&/||
- don't use 'using std::whatever' when 'whatever' is only used once or
  twice.

- Suffix 'u' on numbers to avoid sign comparison warnings
- id() == 0 no longer occurs when removed from a simulation

Replace `typename std::enable_if<...>::type` with
`std::enable_if_t<...>`
Replace various `typename = enable_if_t<...>` with `enable_if_t<..., int> = 0`
which allows for better overloading.

Taking a MemberID allows SharedMember<T>s to be passed, which is no
longer allowed now that the eris_id_t operator in SharedMember<T> is
gone.

`a.transferTo(b)` is almost equivalent to `a.transfer(a, b)` except
that `a` is also cleared (with just `transfer()`, `a` could end up with
neglible amounts for negative quantities when `b` has nearly-identical
positive quantities of the good).

Add a deprecated transferApprox() that forwards its arguments to
transfer().

This makes it easier to write code that is conditionally fixed-size,
such as std::pair<A, B> which is fixed iff both A and B are fixed,
which is more easily determined via a constexpr than SFINAE on member
existance.

This is technically more correct: C++ doesn't guarantee un-namespaced
`size_t`