Jump to content

April 2017 »

232425262728 29
- - - - -

So close, and yet so far

4: Adsense

After implementing a bit of value-based pattern matching support, the Epoch compiler is now down to 7 tests remaining before it's ready for self-hosting.

Of those 7, four are related to templates, which will be the last major endeavor required to get self-hosting viable. Templates are scary and will probably eat up a lot of time.

One of those tests is simply pattern matching not working on built-in functions, notably the "cast" function. I really hate the cast function and want to replace it with a template form instead, but for now it works the way it does, and I'll have to support the current syntax to get the compiler bootstrapped anyways.

Another test is related to value semantics of structures, which has been broken for ages, because of some half-baked optimizations I never finished ironing out. That one may take some time to get right because it requires being very fiddly in how I invoke copies of structures versus handing off references behind the scenes.

The final test in the list is for shadowed variables; unlike most languages, Epoch is actively hostile to identifier shadowing, and forbids it outright. However, the compiler in its current implementation is a bit too zealous with this, and will in fact flag a shadowing error even if you define variables of the same name in totally unrelated scopes. That's kind of annoying and needs to be fixed, but it's low priority.

The three fiddly little tests shouldn't be more than a day apiece to get done. Templates, as I keep observing here, are going to be the nasty heavy lifting bit of the project.

So self hosting remains a visible but still distant target. Another weekend or so should get me down to nothing but templates to fix, and then I can ponder how hideously painful that will be for a while before I drag up the nerve to actually tackle it.

Once self-hosting is done, it's time to take a careful look at what comes next. I'm torn between working on a debugger and working on a profiler, although the two overlap pretty heavily and I might wind up just doing common bits of work for each and seeing what falls out.

Profiling is going to be very important because the new compiler is a few orders of magnitude slower than the C++ one, which is kind of sad but also not entirely unexpected given that I've spent literally zero effort on optimizing the Epoch implementation of the compiler. I can't be bothered waiting more than a second for my programs to compile, so it'll be fun to get back into the performance arena and start hacking away on bottlenecks for a while.

It certainly will be a welcome alternative to all this self-hosting gibberish. I'm seriously bored of watching compiler test suites run...

Sep 29 2013 01:35 AM
Meh. cast() works now.
Sep 29 2013 02:00 AM
Oh, fuck it. I fixed value semantics too, although possibly at the cost of putting a ton of pressure on the GC, we'll have to wait for the profiler to know for sure.

Now the real question is, can I be arsed to fix shadowing before I go to bed? :-P
Sep 29 2013 02:05 AM
Answer: yes, although I did it with a cheap cheesy hack instead of a proper fix, which means that duplicate variable definitions will be silently ignored all over the place. I'll get around to polishing up a proper fix someday™.
Sep 29 2013 06:46 AM


I'll get around to polishing up a proper fix someday™.

What's the TM for? :)

Sep 30 2013 12:13 AM

It's like "someday".

I have no words to describe how much I find your effort admirable. I've given up self-hosting myself.

Note: GameDev.net moderates comments.