Obviously, a 500-fold speed improvement is nothing to sneeze at, and maybe normal people would have looked at this rocket-disguised-as-compiler and decided that, hey, you know, 18.5 milliseconds is good enough.
Well... I have never claimed to be normal.
18.5 milliseconds may be nice, but it isn't there yet. Profiling the compiler shows that a huge amount of time is spent in the parser doing "backtracking." (For reference, I had to move up to a 2MB file for testing, because the 20KB test module I started with doesn't even register on the profiler anymore.)
Backtracking is when a parser looks at a bit of code, says "hmm, I think I'm looking at a Foo" and then gets another byte down the road and goes "whoops! That wasn't a Foo, it's a Foe!" and has to discard all the work it did assuming it was looking at a Foo.
When you do this on a byte-for-byte basis, in a 2MB file, it's clear that things are not as efficient as they could be.
The correct solution to this is lexical analysis. A lexer is a tool which performs lexical analysis. In a nutshell, this is a way of taking the raw code and treating it as a sequence of atomic "tokens" instead of bytes. Now, the parser doesn't think in terms of "F" followed by "o" followed by "o-- whoops, I mean e." Instead, it sees "Foo" and "Foe" as indivisible and obviously different chunks.
During parsing, this dramatically cuts back on backtracking, because the parser can look at the code at a more coarse level. It's like cutting your input size down by a factor roughly equal to the average length of a token in your code - which, as it happens in my Epoch test, is about 6 bytes.
The Epoch compiler is built on top of boost::spirit::qi, which has a companion library boost::spirit::lex. As you can probably guess from the name, lex is a lexer generator. And since, as we've established, 18.5 milliseconds is just too darn slow, it's time to deploy lex.
As a matter of fact, I started this about 24 hours ago, and I've been dabbling in compiler tweaks ever since. It's slow going, because with all the template library magic that's in the Epoch compiler, the build times of that code are getting into the several-minute range. It's endlessly ironic that I'm sacrificing hours of cumulative compile time in the Epoch compiler to shave milliseconds off the runtime... but oh well.
There have been a lot of dumb brain-dead mistakes made in those hours, mainly because I'm splitting my attention between this project and half a dozen other things, so I'm constantly losing track of what I was thinking or doing.
At the moment, I have a hacked version of things running at long last - which is tremendously rewarding. The problem is, it doesn't actually support the Epoch syntax extension mechanism - so all those cool constructs like "if" and "do" and "function" that the standard library provides... don't work.
There seems to be a fundamental assumption someplace in lex that says that you shouldn't ever need to modify the lexing tables on the fly. This strikes me as painfully naive and limiting, so I'm doing my best to hack around it. So far it's slow going, but there is light on the horizon.
I'll poke at this a bit more tonight and then get some sleep and come at it strong tomorrow. Will keep everyone posted!