nim-wiki/Roadmap.md

<table>

<tr>
  <th>Language feature</th>
  <th>Scheduled for</th>
  <th>Notes</th>
</tr>

<tr>
  <td>Implement static[T] properly</td>
  <td>for version 1.0</td>
  <td>The implementation has lots of issues, but what's worse the spec is not nearly clear enough and seems to conflate 2 different ideas.</td>
</tr>

<tr>
  <td>Support overloading of the assignment operator</td>
  <td>for version 1.0</td>
  <td>like <tt class="docutils literal"><span class="pre">fields</span></tt> and destructors an assignment operator needs to be lifted</td>
</tr>

<tr>
  <td>Automatic dereference in more contexts</td>
  <td>might make into 1.0</td>
  <td>important to be able to write more generic containers easily; current plan is rewrite a.f() to f(a[]) but not in further contexts</td>
</tr>

<tr>
  <td>asFunc pragma</td>
  <td>After 1.0</td>
  <td>provides a nice bridge from 'var T' to 'returns T'; can be entirely implemented as a macro with the upcoming types API
</tr>

<tr>
  <td>Extended macros API</td>
  <td>after 1.0</td>
  <td>missing: an API for working with types; the builtin getImpl</td>
</tr>

<tr>
  <td>Pragmas are not supported for 'bindSym'</td>
  <td>after 1.0</td>
  <td>this makes clean macros impossible to write when it comes to pragmas</td>
</tr>

<tr>
  <td>Alias analysis needs to specified and controllable with pragmas</td>
  <td>after 1.0</td>
  <td>alias analysis affects memory safety rules</td>
</tr>

<tr>
  <td>Destructors need to interact with finalizers</td>
  <td>after 1.0</td>
  <td>destructors and finalizers are implemented, but do not interact. <tt class="docutils literal"><span class="pre">new</span></tt> which takes a finalizer needs to be deprecated. People should use a destructor instead. However, destructors should ultimately depend on an escape analysis.</td>
</tr>

<tr>
  <td><tt class="docutils literal"><span class="pre">~</span></tt> Operator for effects</td>
  <td>after 1.0</td>
  <td>it is not entirely clear what <tt class="docutils literal"><span class="pre">~F</span></tt> should mean.</td>
</tr>

<tr>
  <td>Prevent object branch transitions from low(selector)</td>
  <td>after 1.0</td>
  <td>object branch transitions from low(selector) are allowed, because they are simply too useful, but can break memory safety. Now that the language has a proper notion of construction, we can make them safe.</td>
</tr>

<tr>
  <td>Push pragma needs to be reworked</td>
  <td>after 1.0</td>
  <td>it should be distinguished between 'push' for 'type', 'proc' etc.</td>
</tr>

<tr>
  <td>Returning 'var T' is unsafe and needs static analysis to make safe</td>
  <td>after 1.0</td>
  <td>the compiler already checks for trivial examples</td>
</tr>

<tr>
  <td>Some effects like <tt class="docutils literal"><span class="pre">GCMem</span></tt> don't need to be white-listed</td>
  <td>after 1.0</td>
  <td>unclear whether this only affects built-in properties like <tt class="docutils literal"><span class="pre">GCMem</span></tt>, <tt class="docutils literal"><span class="pre">Recursive</span></tt>.</td>
</tr>

<tr>
  <td>Term rewriting macros need more love</td>
  <td>after 1.0</td>
  <td>we need much more tests and use them in production</td>
</tr>

<tr>
  <td>Weaken the requirements for forward declarations</td>
  <td>after 1.0</td>
  <td>two different design ideas exist: one conservative and so will work, the other highly experimental</td>
</tr>

<tr>
  <td>Write tracking</td>
  <td>after 1.0</td>
  <td>algorithm exists. Can be extended to compute lent pointers.</td>
</tr>

</table>

<br>

<table>

<tr>
  <th>Compiler feature</th>
  <th>Scheduled for</th>
  <th>Notes</th>
</tr>

<tr>
  <td>Deterministic code generation</td>
  <td>after 1.0</td>
  <td>deterministic code generation makes C code caching much more effective</td>
</tr>

<tr>
  <td>High level optimizer</td>
  <td>after 1.0</td>
  <td>use the effect system for optimizations; interacts with TR macros; 
      performs:<ul>
      <li>GCSE</li>
      <li>loop invariant hoisting</li>
      <li>escape analysis for string/seq</li>
      <li>inlining of small lambdas</li>
    </ul>
  </td>
</tr>

<tr>
  <td>GPU backend</td>
  <td>after 1.0</td>
</tr>

<tr><td>Pure RC'ing GC</td>
  <td>unknown</td>
  <td>important for better interoperability with Objective-C</td>
</tr>

<tr>
  <td>Symbol files</td>
  <td>after 1.0</td>
  <td>symbol files provide true incremental compilation beyond the caching of generated C code</td>
</tr>

</table>