mhh__

I think the nanopass concept - e.g. lots of well defined IRs passing through a pipeline - is a very good idea for teaching, but I'm interested to see how well it performs and also whether, from the programmers perspective, whether this simplifies code or adds unneeded complexity (Specifically, whether they can be optimized quite as well as a big monolithic compiler).

I'm currently writing a compiler framework - nothing big, as a learning project, and I think I shall try integrate this idea in some way. I feel it would be particularly useful for compiler backends like GCC or LLVM because having well defined pipelines etc. means that one can hook into the framework cleanly (Potentially a huge saving in compile times, e.g. not having to cart around unneeded libraries/symbols).

capnrefsmmat
Chez Scheme [0] is written using the nanopass framework, and it's regarded as one of the fastest Scheme compilers in existence [1]. Before it was rewritten to use the nanopass system, Chez's compiler was known for its performance in terms of lines of code compiled per second; the rewrite slowed it down a bit, but the quality and performance of generated machine code improved. Andy Keep and Kent Dybvig wrote a paper about the project [2]. I haven't browsed the Chez source, but it's a good way to answer your question.

[0] https://github.com/cisco/ChezScheme

[1] http://ecraven.github.io/r7rs-benchmarks/benchmark.html

[2] https://www.cs.indiana.edu/~dyb/pubs/commercial-nanopass.pdf