Last updated: 2019-06-28 15:35:39 +0100

Upstream URL: git clone


View repository

View issue tracker

Contents of follows

Dependency-Injecting Eval for Haskell

This Haskell package is a crude implementation of <code>eval</code>, as found in dynamic languages like Lisp, Python, Javascript, etc. It lets us construct Haskell code programatically (either at run time, or in Template Haskell), and attempt to evaluate it.

What sets this package apart from other <code>eval</code> implementations, is the ability to control which packages and modules are available during evaluation. This is achieved by calling out to the Nix package manager.

Implementation Details

<code>Expr</code> is the type of expressions, which contains a list of package names, a list of modules to import, a list of compiler flags, a list of <code>String</code>s to put in the generated module and a <code>String</code> of Haskell code to evaluate. All of these are just <code>String</code>s internally, but we use wrappers to prevent accidentally using packages as modules, etc.

A few combinators are provided for common manipulations, for example <code>qualified "Foo" "bar"</code> will produce the expression <code>""</code> with <code>"Foo"</code> in its module list. The <code>OverloadedStrings</code> extension allows packages, modules, flags and expressions to be written as literals. Note that literal expressions are given an empty context; you will have to specify any required modules, packages, etc. separately.

When evaluated, the Haskell code is prefixed by an import of each module, the “preamble” strings (if any) and wrapped in <code>main = putStr (..)</code>. This code is piped into <code>runhaskell</code>. If any flags are specified, they are appended as arguments to the <code>runhaskell</code> command.

The <code>runhaskell</code> process itself is invoked via the <code>nix-shell</code> command, which provides all of the required packages via the <code>ghcWithPackages</code> mechanism of nixpkgs. Packages are taken from nixpkgs’s <code>haskellPackages</code> set by default, which can be overridden by setting the <code>NIX_EVAL_HASKELL_PKGS</code> environment variable to the path of a Nix file. Note that the package names used in your Haskell code should correspond to the keys in this package set, which might differ from those used on Hackage.

If the process exits successfully, its stdout will be returned wrapped in <code>Just</code>; otherwise <code>Nothing</code> is returned. If you wish to alter the <code>main</code> implementation, use <code>Language.Eval.Internal.eval'</code>

This implementation is a little rough; for example, you may prefer to use <code>Text</code> rather than <code>String</code>; use a better representation like the syntax trees from TemplateHaskell or <code>haskell-src-exts</code> instead; or accumulate packages and modules monadically.

The intention of this library is to provide a simple, minimal base to support such design choices, and <code>String</code> is the lowest common denominator. You’re welcome, and encouraged, to build more sophisticated APIs; as long as you can pretty-print to a <code>String</code>, they should work out of the box.

This is also why we return the contents of stdout, rather than trying to parse it into a more appropriate type: it’s not our place to choose how the result should be parsed, so we avoid the problem; by that point, our job is done.


<ul> <li>Since evaluation takes place in a separate GHC process, there can be no sharing of data outside the strings provided (unless you provide a separate mechanism like a FIFO)</li> <li>Expressions are wrapped in <code>putStr</code>, so the expression must be a <code>String</code>. You may need to marshall your data into a form which is more amenable to serialising/deserialising via <code>String</code>.</li> <li>Evaluation is SLOW! More specifically, <code>eval</code> has a very high latency, so it’s <em>much</em> more efficient to <code>eval</code> one big collection of values than it is to <code>eval</code> each individually.</li> <li>Evaluation time is highly variable, since the required packages may need to be compiled. Nix caches build products, so subsequent calls using the same packages will be quicker; however, my machine still takes about 2 seconds to instantiate a cached environment.</li> <li>Output is captured from stdout, so if your expression triggers side-effects they’ll appear in your result (this may be desirable, but keep it in mind).</li> <li>Evaluation doesn’t always compose, ie. just because <code>x</code> and <code>y</code> evaluate successfully doesn’t mean that some combination of them will. Obviously an ill-typed combination will fail, but other reasons include: <ul> <li>Combining both import lists can make names ambiguous. For this reason you should always try to qualify your expressions.</li> <li>Global properties may conflict between modules, like overlapping typeclass instances.</li> <li>Combining both package lists can make modules ambiguous.</li> <li>If the dependencies of two packages conflict, evaluation will fail.</li> </ul></li> <li>As with any kind of <code>eval</code>, there is absolutely no security. Do not pass potentially-malicious user input to this library! Not only can arbitrary Haskell code be run (eg. using <code>unsafePerformIO</code>, but the flags are also a shell injection vector.</li> </ul>