when i set out to find better program semantics, one of the background requirements was "make it canonicalizable", and then i totally forgot about it and just went by instinct.
now i realize nudl programs fit the requirement: the program is formulated as a non-recursive dataflow DAG, with all mutations happening on weak backedges. and, because there is a mechanism to encapsulate mutations as storage operations, each iteration of the program can be made pure.
how this is good:
* two DAGs are structurally identical if and only if their canonical forms match. this enables fast hashing, cache lookups, and deduplication.
* structural diffs become tractable. you can detect which nodes/edges changed between two versions of the program.
* unambiguous serialization: diffing then makes it possible to restore most of the state under schema change.
@lritter Nice! Very vague memories of wondering about "canonicalisability" maybe allowing for greater sophistication in editing / navigating a language.
Sadly not recalling much more than that atm, but I think part of it was the idea of the language always being some of the way through the compilation process – tokenisation, at least – while we're writing it.
@miblo ah. this particular application would be a bit difficult here since the program allows relating data completely out of order - that's a major feature, in fact.
however, once parsed, the compiler offers many avenues for parallelization, primarily because the language itself was designed for concurrency.
(it's a side effect of specialization - if your specialized code is easy to parallelize, then any interpreter pass (including typechecking) can be, too)
@lritter Ah right! It's coming back to me a bit, pretty sure I'd've wanted it to allow out-of-order declarations, and I'm recalling this weird notion of having the ordering itself be canonicalised – based on order of access / execution, dependency on prior computations – whose "rendered position in the file" is represented as an offset from the canonical position according to where the programmer wrote it.
Dunno, maybe that couldn't've worked after all…
LR
Matt Mascarenhas