Re:Building Fast Interpreters in Rust
by saagarjha on Monday March 04, 2019 @12:41PM
(#19303234)
A better way to get good performance is to thread your switch statement, which is hard to do explicitly in Rust last time I tried (maybe you could do this if you mark functions as inlinable?).
[ Parent
| on HN ]
With current branch predictors threaded code might not make a big difference like it used to.
What do you mean by "thread your switch statement"?
Re:Building Fast Interpreters in Rust
by ridiculous_fish on Monday March 04, 2019 @01:27PM
(#19303780)
The "big switch statement" approach is for each bytecode instruction to complete by jumping to a centralized dispatch location (i.e. the switch statement).
The "threaded" approach is for each bytecode instruction to complete by decoding and jumping to the handler for the next instruction.
Basically instead of "break" you have `goto handlers[nextIp->opcode].`
The advantages of threading are fewer jumps and better branch prediction (since branch prediction is tied to IP). The disadvantages are slightly larger code and compilers struggle to optimize it, since the control flow is not structured.
Re:Building Fast Interpreters in Rust
by filereaper on Monday March 04, 2019 @06:55PM
(#19306890)
This method of design is called a Continuation Passing Style Interpreter. [1]
Here's a production version from OpenJ9's JVM ByteCode Interpreter. [2]
[1] https://kseo.github.io/posts/2017-01-09-continuation-passing...
[2] https://github.com/eclipse/openj9/blob/01be53f659a8190959c16...
I think he means have a different thread for each case or a subset of cases instead of having any explicit switch statement.