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.
I work on PyTorch torch.compile and it’s a tracing compiler as well. Perhaps this domain is very narrow though; it is also a very not conventional compiler, you’d probably be deeply offended by some of the stuff we do! ;)
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.
ML compilers in particular go beyond even the level of stability you would expect from numerical programs. Due to how the SIMT model of thread/warp divergence works, the hardware heavily punishes unstable branches. E.g. if you have 32 threads taking a branch then recoalescing on a barrier -- if they all go the same direction then they can go down the execution pipe as a single bundle, but if 1 takes it while 31 don't, then that's 2x the ex-pipe usage by default (and if you have e.g. a computed-branch, performance goes out the window). Consequently, the whole stack is built around the expectation of stable control flow, even to the detriment of performance (from a local perspective).
ML frameworks even take advantage of this to compute, ahead-of-time, how much memory will be used at different points in the program graph, and thereafter schedule memcpy's to make space as necessary. Of course this only works for well-behaved program classes, but e.g. most LLM architectures fit into that category. Interestingly MoE models don't, since they require data-dependent control flow, thus the recent push towards accommodating dynamism in frameworks (like JAX, which until ~recently couldn't handle it at all).
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.
I am actively opposed to this design for a first compiler. There is no need for a lexer with a recursive descent parser. Register allocation is also an unnecessary distraction. It is better for a first compiler to compile to a higher level language in which neither register assignment nor memory management are necessary.
Optimizing compilers are suboptimal in that they waste enormous amount of time optimizing code that can't or needn't be optimized and even where the optimizations are helpful, they are opaque and at risk of unexpectedly regressing both due to small changes at the source code level or changes in the compiler optimizer, both of which are quite insidious.
If instead of optimizing compilers, we had languages that allowed for seamless interop between low level and high level functions, then perhaps an llm becomes the optimizer (or you can invoke the compiler to optimize a specific function by source level rewrite). The benefit of this compared to a traditional optimizing compiler is that the optimization is done once per function and never repeated (until prompted) and the implementation is human readable and not buried in a binary. Moreover, and perhaps even more importantly, by not doing optimizations in the compiler, compilation times can be much faster, easily 100-1000x than state of the art optimizing compiler, while generating equivalent or even better runtime performance. As it has been said: premature optimization is the root of all evil.
The academic literature on register allocation is scary.
First is presented a linear time optimal algorithm for graph coloring then it is claimed better can be done by a O(N^2) algorithm that uses a heuristic.
I do believe the dragon book got caught with the emperor's new register allocator and the literature hasn't really recovered yet.
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
I've also never heard this before. The closest I can think of is fuzzing test suites used to find panics in various crates, but that's neither machine learning nor in the compiler itself.
I know that datalog is used for the borrow checking logic, so I could also maybe imagine someone describing something like that in some hand wavy way like "proof by machine to detect up front whether the program is safe or might crash" and that getting misinterpreted, but that seems like a stretch
I'm a fan of the structure used in Essentials of Compilation [1][2] and Writing a C Compiler [3]. If you want to start with interpreters, I like Essentials of Programming Languages [4]. I have to admit, as popular as Crafting Interpreters is on this site and others, I'm not a fan. Others seem to love it so it's worth a shot, and freely available.
EOC and EOPL are a bit on the academic side, but, I think, they're highly approachable aside from the issues some people have with Scheme and Racket (the Python version of EOC would address that issue). Afterwards, I think the other, deeper and more academic texts on compilers become more approachable.
I'd say check out Crafting Interpreters [1]. It has 2 parts, the first in Java for doing a treewalk Interpreter in Java before going farther with a version written in C.
Not GP, but after doing Crafting Interpreters I was kinda left with a gap in my knowledge regarding the conversion of an AST into native code. Also kinda missing was optimization passes over an AST. I somewhat understand the idea, but it would definitely be nice to have a more guided book/article for this.
Crafting Interpreters is definitely a recommended read, but it stops at Interpreters (fair enough, the book is thick enough). Crafting Compilers would need at least 4-5 extra chapters IMO.
Previously...
CS 6120: Advanced Compilers: The Self-Guided Online Course - https://news.ycombinator.com/item?id=39577878 - March 2024 (102 comments)
Advanced Compilers: Self-Guided Online Course - https://news.ycombinator.com/item?id=35130975 - March 2023 (82 comments)
Advanced Compilers: Self-Guided Online Course - https://news.ycombinator.com/item?id=25386756 - Dec 2020 (232 comments)
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.
I work on PyTorch torch.compile and it’s a tracing compiler as well. Perhaps this domain is very narrow though; it is also a very not conventional compiler, you’d probably be deeply offended by some of the stuff we do! ;)
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.
ML compilers in particular go beyond even the level of stability you would expect from numerical programs. Due to how the SIMT model of thread/warp divergence works, the hardware heavily punishes unstable branches. E.g. if you have 32 threads taking a branch then recoalescing on a barrier -- if they all go the same direction then they can go down the execution pipe as a single bundle, but if 1 takes it while 31 don't, then that's 2x the ex-pipe usage by default (and if you have e.g. a computed-branch, performance goes out the window). Consequently, the whole stack is built around the expectation of stable control flow, even to the detriment of performance (from a local perspective).
ML frameworks even take advantage of this to compute, ahead-of-time, how much memory will be used at different points in the program graph, and thereafter schedule memcpy's to make space as necessary. Of course this only works for well-behaved program classes, but e.g. most LLM architectures fit into that category. Interestingly MoE models don't, since they require data-dependent control flow, thus the recent push towards accommodating dynamism in frameworks (like JAX, which until ~recently couldn't handle it at all).
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.
Has LuaJIT been superseded?
Got any other recommended resources on building compilers?
PyTorch?
How does this compare to Nora Sandler's "Writing a C compiler" in terms of the potential gains for the reader?
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.
I am actively opposed to this design for a first compiler. There is no need for a lexer with a recursive descent parser. Register allocation is also an unnecessary distraction. It is better for a first compiler to compile to a higher level language in which neither register assignment nor memory management are necessary.
Optimizing compilers are suboptimal in that they waste enormous amount of time optimizing code that can't or needn't be optimized and even where the optimizations are helpful, they are opaque and at risk of unexpectedly regressing both due to small changes at the source code level or changes in the compiler optimizer, both of which are quite insidious.
If instead of optimizing compilers, we had languages that allowed for seamless interop between low level and high level functions, then perhaps an llm becomes the optimizer (or you can invoke the compiler to optimize a specific function by source level rewrite). The benefit of this compared to a traditional optimizing compiler is that the optimization is done once per function and never repeated (until prompted) and the implementation is human readable and not buried in a binary. Moreover, and perhaps even more importantly, by not doing optimizations in the compiler, compilation times can be much faster, easily 100-1000x than state of the art optimizing compiler, while generating equivalent or even better runtime performance. As it has been said: premature optimization is the root of all evil.
The academic literature on register allocation is scary.
First is presented a linear time optimal algorithm for graph coloring then it is claimed better can be done by a O(N^2) algorithm that uses a heuristic.
I do believe the dragon book got caught with the emperor's new register allocator and the literature hasn't really recovered yet.
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.
Well there is advanced and there is advanced.
This course is just a second course on compilers for people who had an introduction. And a great one at that.
A good modern, practical and decently optimizing compiler can do just fine without all the things you've mentioned, including vectorization.
Besides, most programming language implementations never go beyond the basic SSA toolkit.
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've also never heard this before. The closest I can think of is fuzzing test suites used to find panics in various crates, but that's neither machine learning nor in the compiler itself.
I know that datalog is used for the borrow checking logic, so I could also maybe imagine someone describing something like that in some hand wavy way like "proof by machine to detect up front whether the program is safe or might crash" and that getting misinterpreted, but that seems like a stretch
Is there also a self guided course for "basic compilers", before stepping into an advanced level?
I'm a fan of the structure used in Essentials of Compilation [1][2] and Writing a C Compiler [3]. If you want to start with interpreters, I like Essentials of Programming Languages [4]. I have to admit, as popular as Crafting Interpreters is on this site and others, I'm not a fan. Others seem to love it so it's worth a shot, and freely available.
EOC and EOPL are a bit on the academic side, but, I think, they're highly approachable aside from the issues some people have with Scheme and Racket (the Python version of EOC would address that issue). Afterwards, I think the other, deeper and more academic texts on compilers become more approachable.
[1] https://mitpress.mit.edu/9780262047760/essentials-of-compila... - Racket version, has an open access version
[2] https://mitpress.mit.edu/9780262048248/essentials-of-compila... - Python version, has an open access version
[3] https://nostarch.com/writing-c-compiler - Your choice of implementation language
[4] https://mitpress.mit.edu/9780262062794/essentials-of-program... - Scheme, but works in Racket
Haven't tried it myself, but there is "the-super-tiny-compiler": https://github.com/jamiebuilds/the-super-tiny-compiler
I'd say check out Crafting Interpreters [1]. It has 2 parts, the first in Java for doing a treewalk Interpreter in Java before going farther with a version written in C.
1. https://craftinginterpreters.com/
Not GP, but after doing Crafting Interpreters I was kinda left with a gap in my knowledge regarding the conversion of an AST into native code. Also kinda missing was optimization passes over an AST. I somewhat understand the idea, but it would definitely be nice to have a more guided book/article for this.
Crafting Interpreters is definitely a recommended read, but it stops at Interpreters (fair enough, the book is thick enough). Crafting Compilers would need at least 4-5 extra chapters IMO.
Hmm I would have hoped for something more formal and that's focused on compiled runtimes, instead of dynamic runtimes
Still, I appreciate you replying, I'm sure you meant to be helpful!
I'm super curious what alexia massalin is up to these days, besides collecting microunity patent royalties