The section on dynamic compilers is more or less all about trace compilation. Generally, trace compilation is a dead end and has been abandoned repeatedly. The more important concepts here are type feedback and speculation and deoptimization, as well as making fast compilers and tiering.
The course overall looks good, and it's great that so much is available online, so well done, Adrian.
Thanks, Ben. I admit I mostly think tracing is just a mind-expanding concept to learn about, even if history has proven it’s not very practical as an organizing principle. But as you say, I’d love to offer more context on “what actually seems to work” industrially.
Yeah, it is conceptually interesting. I like giving students perspective and in 770 (https://www.cs.cmu.edu/~wasm/cs17-770/fall2025/) I might spend half or less of a lecture on tracing and I don't pull punches on how I think it ends up not really working well in real systems. It's a good opportunity to talk about program behavior and the cost/benefit of speculation.
We spend a lot more time on type feedback, ICs, and deoptimization which are the more universal concepts that can be applied to multiple different compiler designs.
Interesting. I think numerical computing is a narrow enough domain where programs have very well-behaved control flow, which avoids most of the problems of trace compilation. Loops over branchy code, which are really common in general programs, are very difficult to make work well with tracing.
Numerical programs being very stable in terms of control is what enables GPU parallelization and loop optimizations in the long tradition of Fortran compilers. Optimizations like loop tiling, interchange, strip mining, etc aren't going to be easy to do with trace compilation.
Anyway my comment was more directed toward trace compilation in the context of dynamic languages, and there I think it's pretty well established it only works well for small programs.
The TraceMonkey paper was on my qual reading list, and my quals happened to be around the time TraceMonkey was ripped out of SpiderMonkey. I was talking with one of the developers at the time (I think it was Jason Orendorff?), who said that tracing just doesn't work out, but there was limited circumstances where he thought it might work out... but I've completely forgotten what those circumstances were.
I'm a bit confused about what makes this course "advanced." Most of the topics (dead code elimination, data flow, dominator analysis, SSA form) seem like they belong in a first course on compilers.
The short answer is that compilers is basically broken up into two courses, with the first course largely being the minimum necessary to build a compiler (lexing, parsing, codegen, register allocation), and the second course being how to build an optimizing compiler.
Looks like there is quite a bit of overlap with regards to the optimizing parts between these two courses. I guess it's switching from the dragon book to academic papers that makes it advanced.
I think a lot of the non-professionals start with parsing and do not get exposed to backend. I have read two books about interpreters/compilers and they don't touch the backend very much.
That's part of it. I think another part is that it seems like the students are asked to read the papers behind a lot of the concepts, and academic literature is not generally very accessible to undergrads. (Not that they can't read it, but without someone guiding you through at least the first few papers, it can be a frustrating experience for many.)
Saw a podcast that talked about the rust compiler, which apparently included machine learning algorithms at some points to determine whether or not you had code that could crash your system
The section on dynamic compilers is more or less all about trace compilation. Generally, trace compilation is a dead end and has been abandoned repeatedly. The more important concepts here are type feedback and speculation and deoptimization, as well as making fast compilers and tiering.
The course overall looks good, and it's great that so much is available online, so well done, Adrian.
Thanks, Ben. I admit I mostly think tracing is just a mind-expanding concept to learn about, even if history has proven it’s not very practical as an organizing principle. But as you say, I’d love to offer more context on “what actually seems to work” industrially.
Yeah, it is conceptually interesting. I like giving students perspective and in 770 (https://www.cs.cmu.edu/~wasm/cs17-770/fall2025/) I might spend half or less of a lecture on tracing and I don't pull punches on how I think it ends up not really working well in real systems. It's a good opportunity to talk about program behavior and the cost/benefit of speculation.
We spend a lot more time on type feedback, ICs, and deoptimization which are the more universal concepts that can be applied to multiple different compiler designs.
> Generally, trace compilation is a dead end and has been abandoned repeatedly.
JAX is a tracing compiler!
(I know, I know, it sits in an extremely different part of the problem space than TraceMonkey or LuaJIT. Still.)
Interesting. I think numerical computing is a narrow enough domain where programs have very well-behaved control flow, which avoids most of the problems of trace compilation. Loops over branchy code, which are really common in general programs, are very difficult to make work well with tracing.
Numerical programs being very stable in terms of control is what enables GPU parallelization and loop optimizations in the long tradition of Fortran compilers. Optimizations like loop tiling, interchange, strip mining, etc aren't going to be easy to do with trace compilation.
Anyway my comment was more directed toward trace compilation in the context of dynamic languages, and there I think it's pretty well established it only works well for small programs.
and PyPy right?
The TraceMonkey paper was on my qual reading list, and my quals happened to be around the time TraceMonkey was ripped out of SpiderMonkey. I was talking with one of the developers at the time (I think it was Jason Orendorff?), who said that tracing just doesn't work out, but there was limited circumstances where he thought it might work out... but I've completely forgotten what those circumstances were.
PyTorch?
Got any other recommended resources on building compilers?
Has LuaJIT been superseded?
I'm a bit confused about what makes this course "advanced." Most of the topics (dead code elimination, data flow, dominator analysis, SSA form) seem like they belong in a first course on compilers.
Well, course numbers are regular enough that you can look up what the "intro compilers" course is: https://www.cs.cornell.edu/courses/cs4120/2026sp/?schedule
The short answer is that compilers is basically broken up into two courses, with the first course largely being the minimum necessary to build a compiler (lexing, parsing, codegen, register allocation), and the second course being how to build an optimizing compiler.
Looks like there is quite a bit of overlap with regards to the optimizing parts between these two courses. I guess it's switching from the dragon book to academic papers that makes it advanced.
I have read TONS of material about it*, and none of that is part of the majority of that!
In fact, the "backend" be compiler or interpreter is nearly always left as "exercise to reader".
You can't imagine how much is left to be discovered, from how make a closure, track environment, do pattern matching, memory representation, etc.
EVERYTHING interesting is something you need to look for.
P.D: This only one of the years:https://gist.githubusercontent.com/mamcx/e1743571b9a1ea163a7...
I think a lot of the non-professionals start with parsing and do not get exposed to backend. I have read two books about interpreters/compilers and they don't touch the backend very much.
Maybe this is introductory for backend?
That's part of it. I think another part is that it seems like the students are asked to read the papers behind a lot of the concepts, and academic literature is not generally very accessible to undergrads. (Not that they can't read it, but without someone guiding you through at least the first few papers, it can be a frustrating experience for many.)
What is advanced then? Good coverage of dce, data flow, ssa, intruction selection and reg alloc is actually like 98% of the backend.
Perhaps polyhedral optimization, tiling, scalar evolution, vectorization...
I guess garbage collection is pretty advanced in the syllabus.
Saw a podcast that talked about the rust compiler, which apparently included machine learning algorithms at some points to determine whether or not you had code that could crash your system
I've never heard about that and I'm pretty sure it's incorrect (although "machine learning" is a wide term), do you have a source for that?
I'm super curious what alexia massalin is up to these days, besides collecting microunity patent royalties