[<<][meta][>>][..]
Sun Feb 14 13:41:51 CET 2010

Environment is only necessary during threading

Let's rephrase...  What is important?  When flattening an expression
tree/graph, it needs to be determined which nodes are the same.

What should an environment do?

  - provide an ANF/SSA view of expression graph

Anything else is implementation.  In first approximation this means:
an environment should keep track of sharing based on some form of
equivalence, i.e. object equivalence (as a consequence of data
sharing at the meta level) or other equivalences.

So basically, we never need to _merge_ environments, as there is only
one instance that is threaded through the flattening.  It looks like
the problem is simpler this way.

Ok, this works pretty well.

  - given terms, compute a fold over them to accumulate the sharing
    environment, and replace all terms with shared variants.

  - provide conversion of environment to node-naming function (for
    obtaining register names)

  - provide fold over nodes in environment

This then gives a way to convert a term (with sharing and possibly
other equivalence relation) to a flat ANF-style structure.




[Reply][About]
[<<][meta][>>][..]