c3lang
Speed up parsing using triehash.
I am not sure if it will improve the symbol lookup that much, but if you want you can try it out.
Yes, it's a half-step. My ideas:
- replace fnv2 to faster hasher. According to SMhasher I'd choose komihash or pearsonB
- replace hash table to rax or libart. rax is well tested on zillions of Redis servers :)
- in the long term to use ctl
One thing to understand is that profiling shows that the cost when loading the hash map a lot (lots of symbols) what costs time is mainly memory allocation. This is despite using a mem map implementation. When you request memory, it isn't really allocated until it's touched. And doing that touch is sufficiently slow that it's better to actually go through and touch a bunch of pages in a row than to wait until they're needed(!)
Anyway, the cost here is mainly due to the usage pattern, and a "faster hash map" doesn't really matter other than how the allocation pattern works. So essentially any off-the-shelf hash map will have worse tuning abilities than a custom one.
Yes fnv1a might not be fastest, but it's not the bottleneck and it's dead simple. The problem in terms of compile time will remain LLVM, clocking in at 1-2 magnitudes more expensive than anything else. Also, in order to tune things a much larger corpus of code is needed than even the standard library. I would like to see something in the order of 500.000 kloc to test on. And we don't have that yet.
Hashing is not showing up anywhere, so changing hash is not going to help. What we need is a faster LLVM...