Huh. This is a very... "interesting" application for an LLM. I'm not the brightest crayon in the box, but if anyone else would like to follow along with my non-expert opinion as I read through the paper, here's my take on it.

It's pretty important for compilers / decompilers to be reliable and accurate -- compilers behaving in a deterministic and predictable way is an important fundamental of pipelines.

LLMs are inherently unpredictable, and so using an LLM for compilation / decompilation -- even an LLM that has 99.99% accuracy -- feels a bit odd to include as a piece in my build pipeline.

That said, let's look at the paper and see what they did.

They essentially started with CodeLlama, and then went further to train the model on three tasks -- one primary, and two downstream.

The first task is compilation: given input code and a set of compiler flags, can we predict the output assembly? Given the inability to verify correctness without using a traditional compiler, this feels like it's of limited use on its own. However, training a model on this as a primary task enables a couple of downstream tasks. Namely:

The second task (and first downstream task) is compiler flag prediction / optimization to predict / optimize for smaller assembly sizes. It's a bit disappointing that they only seem to be able to optimize for assembly size (and not execution speed), but it's not without its uses. Because the output of this task (compiler flags) are then passed to a deterministic function (a traditional compiler), then the instability of the LLM is mitigated.

The third task (second downstream task) is decompilation. This is not the first time that LLMs have been trained to do better decompilation -- however, because of the pretraining that they did on the primary task, they feel that this provides some advantages over previous approaches. Sadly, they only compare LLM Compiler to Code Llama and GPT-4 Turbo, and not against any other LLMs fine-tuned for the decompilation task, so it's difficult to see in context how much better their approach is.

Regarding the verifiability of the disassembly approach, the authors note that there are issues regarding correctness. So the authors employ round-tripping -- recompiling the decompiled code (using the same compiler flags) to verify correctness / exact-match. This still puts accuracy in the 45% or so (if I understand their output numbers), so it's not entirely trustworthy yet, but it might be able to still be useful (especially if used alongside a traditional decompiler, and this model's outputs only used when they are verifiably correct).

Overall I'm happy to see this model be released as it seems like an interesting use-case. I may need to read more, but at first blush I'm not immediately excited by the possibilities that this unlocks. Most of all, I would like to see it explored if these methods could be extended to optimize for performance -- not just size of assembly.

It is normally not a necessary feature of a compiler to be determistic. A compiler should be correct against a specification. If the specification allows indeterminism a compiler should be able to exploit them. I remember the story of the sather-k compiler that did things differently based on the phase of the moon.

It's technically correct that a language specification is rarely precise enough to require compiler output to be deterministic.

But it's pragmatically true that engineers will want to murder you if your compiler is non-deterministic. All sorts of build systems, benchmark harnesses, supply chain validation tools, and other bits of surrounding ecosystem will shit the bed if the compiler doesn't produce bitwise identical output on the same input and compiler flags.

Can vouch for this having fixed non-determinism bugs in a compiler. Nobody is happy if your builds aren't reproducible. You'll also suffer crazy performance problems as everything downstream rebuilds randomly and all your build caches randomly miss.

NixOS with its nixpkgs [0] and cache [1] would also not work if compilers weren't reproducible. Though they won't use something like PGO or some specific optimization flags as these would very likely lead to unreproducible builds. For example most distros ship a PGO optimized build of Python while nixos does not.

[0] https://github.com/nixos/nixpkgs

[1] https://cache.nixos.org/

PGO can be used in such situations, but the profile needs to be checked in. Same code + same profile -> same binary (assuming the compiler is deterministic, which is tested quite extensively).

There are several big projects that use PGO (like Chrome), and you can get a deterministic build at whatever revision using PGO as the profiles are checked in to the repository.

It’s called autofdo although I’ve struggled to get it working well in Rust.

It's not called AutoFDO. AutoFDO refers to a specific sampling-based profile technique out of Google (https://dl.acm.org/doi/abs/10.1145/2854038.2854044). Sometimes people will refer to that as PGO though (with PGO and FDO being somewhat synonymous, but PGO seeming to be the preferred term in the open source LLVM world). Chrome specifically uses instrumented PGO which is very much not AutoFDO.

PGO works just fine in Rust and has support built into the compiler (https://doc.rust-lang.org/rustc/profile-guided-optimization....).

I wasn’t trying to conflate the two. PGO traditionally meant a trace build but as a term it’s pretty generic, at least to me to the general concept of “you have profile information that replaces generically tuned heuristics that the compiler uses). AutoFDO I’d classify as an extension to that concept to a more general PGO technique; kind of ThinLTO vs LTO. Specifically, it generates the “same” information to supplant compiler heuristics, but is more flexible in that the sample can be fed back into “arbitrary” versions of the code using normal sampling techniques instead of an instrumented trace. The reason sampling is better is that it more easily fits into capturing data from production which is much harder to accomplish for the tracing variant (due to perf overheads). Additionally, because it works across versions the amortized compile cost drops from 2x to 1x because you only need to reseed your profile data periodically.

I was under the impression they had switched to AutoFDO across the board but maybe that’s just for their cloud stuff and Chrome continues to run a representative workload since that path is more mature. I would guess that if it’s not being used already, they’re exploring how to make Chrome run AutoFDO for the same reason everyone started using ThinLTO - it brought most of the advantages while fixing the disadvantages that hampered adoption.

And yes, while PGO is available natively, AutoFDO isn’t quite as smooth.

I'm not sure where you're getting your information from.

Chrome (and many other performance-critical workloads) is using instrumented PGO because it gives better performance gains, not because it's a more mature path. AutoFDO is only used in situations where collecting data with an instrumented build is difficult.

Last I looked AutoFDO builds were similar in performance to PGO as ThinLTO vs LTO is. I’d say that collecting data with an instrumented Chrome build is extremely difficult - you’re relying on your synthetic benchmark environment which is very very different from the real world (eg extensions aren’t installed, the patterns of websites being browsed is not realistic, etc). There’s also a 2x compile cost because you have to build Chrome twice in the exact same way + you have to run a synthetic benchmark on each build to generate the trace.

I’m just using an educated guess to say that at some point in the future Chrome will switch to AutoFDO, potentially using traces harvested from end user computers (potentially just from their employees even to avoid privacy complaints).

You can make the synthetic benchmarks relatively accurate, it just takes effort. The compile-time hit and additional effort is often worth it for the extra couple percent for important applications.

Performance is also pretty different on the scales that performance engineers are interested in for these sorts of production codes, but without the build system scalability problems that LTO has. The original AutoFDO paper shows an improvement of 10.5%->12.5% going from AutoFDO to instrumented PGO. That is pretty big. It's probably even bigger with newer instrumentation based techniques like CSPGO.

They also mention the exact reasons that AutoFDO will not perform as well, with issues in debug info and losing profile accuracy due to sampling inaccuracy.

I couldn't find any numbers for Chrome, but I am reasonably certain that they have tried both and continue to use instrumented PGO for the extra couple percent. There are other pieces of the Chrome ecosystem (specifically the ChromeOS kernel) that are already optimized using sampling-based profiling. It's been a while since I last talked to the Chromium toolchain people about this though. I also remember hearing them benchmark FEPGO vs IRPGO at some point and concluding that IRPGO was better.

Yeah, and nixpkgs also, last time I checked, does patch GCC/ clang to ensure determinism. Many compilers and toolchain by default want to, e.g., embed build information that may leak from the build env in a non-deterministic/ non-reprodicible manner.

Yup. Even so much as inserting the build timestamp into the generated executable (which is strangely common) causes havoc with build caching.

Plus, these models are entirely black boxes. Even given weights, we don't know how to look at them and meaningfully tell what's happening - and not only that, but training these models is likely not cheap at all.

Stable output is how we can verify that attacks like the one described in Reflections on Trusting Trust[0] don't happen.

[0] https://www.cs.cmu.edu/~rdriley/487/papers/Thompson_1984_Ref...

NVCC CUDA builds were nondeterministic last time I checked, it made certain things (trying to get very clever with generating patches) difficult. This was also hampered by certain libraries (maybe GTSAM?) wanting to write __DATE__ somewhere in the build output, creating endlessly changing builds.

In parallel computing you run into nondeterminism pretty quickly anyways - especially with CUDA because of undetermined execution order and floating point accuracy.

Yes, at runtime. Compiling CUDA doesn’t require a GPU, though, and doesn’t really use “parallel computing”. I think CUDA via clang gets this right and will produce the same build every time - it was purely an NVCC issue.

LLMs can be deterministic if you set the random seed and pin it to a certain version of the weights.

My bigger concern would be bugs in the machine code would be very, very difficult to track down.

Just fix the random seed :)

I’m amused by the possibility of a compiler having a flag to set a random seed. (with a fixed default, of course).

If you hit a compiler bug, you could try a different seed to see what happens.

Or how about a code formatter with a random seed?

Tool developers could run unit tests with a different seed until they find a bug - or hide the problem by finding a lucky seed for which you have no provable bugs :)

> But it's pragmatically true that engineers will want to murder you if your compiler is non-deterministic. All sorts of build systems, benchmark harnesses, supply chain validation tools, and other bits of surrounding ecosystem will shit the bed if the compiler doesn't produce bitwise identical output on the same input and compiler flags.

I think that’s rather true nowadays, but hasn’t always been thus. Back in the 20th century, non-deterministic compiler output was very common - even if only due to the common practice of embedding the compilation timestamp in the resulting executable - and very few ever cared. Whereas nowadays, there is a much bigger culture of hermetic/reproducible build processes, in which stuff like embedding compilation timestamps in the executable or object files is viewed as an antipattern.

> It is normally not a necessary feature of a compiler to be determistic. A compiler should be correct against a specification.

That sounds like a nightmare. Optimizing code to play nice with black-box heuristic compilers like V8's TurboFan is, already in fact, a continual maintenance nightmare.

If you don't care about performance, non-deterministic compilation is probably "good enough." See TurboFan.

LLMs are deterministic. We inject randomness after the fact, just because we don't like our text being deterministic. Turn temperature to 0 and you're good.

But temperature 0 LLM's don't exhibit the emergent phenomena we like, even in apparently non-creative tasks. The randomness is, in some sense, a cheap proxy for an infeasible search over all completion sequences, much like simulated annealing with zero temperature is a search for a local optimum but adding randomness makes it explore globally and find more interesting possibilities.

Sure but you could add pseudo random noise instead and get the same behavior while retaining determinism.

Temperature is at ~1.2 in this thread, here's some 0.0:

- Yes, temperature 0.0 is less creative.

- Injecting pseudo-random noise to get deterministic creative outputs is "not even wrong", in the Wolfgang Pauli sense. It's fixing something that isn't broken, with something that can't fix it, that if it could, would be replicating the original behavior - more simply, it's proposing non-deterministic determinism.

- Temperature 0.0, in practice, is an LLM. There aren't emergent phenomena, in the sense "emergent phenomena" is used with LLMs, missing. Many, many, many, applications use this.

- In simplistic scenarios, on very small models, 0.0 could get stuck literally repeating the same token.

- There's a whole other layer of ex. repeat penalties/frequency penalties and such that are used during inference to limit this. Only OpenAI and llama.cpp expose repeat/frequency.

- Temperature 0.0 is still non-deterministic on ex. OpenAI, though substantially the same, and even the same most of the time. It's hard to notice differences. (Reproducible builds require extra engineering effort, the same way ensuring temperature = 0.0 is truly deterministic requires engineering effort.)

- Pedantically, only temperature 0.0 at the same seed (initial state) is deterministic.

Even then, though, the output could change drastically based on a single change to the input (such as a comment).

That's not something you want in a compiler.

It is very important for a compiler to be deterministic. Otherwise you can't validate the integrity of binaries! We already have issues with reproducibility without adding this shit in the mix.

Reproducible builds are an edge case, that require determistic compilation for sure. But profile based optimisation or linker address randomisation are sometimes also useful. While rule out one thing for the other. Normally you can easily turn on and off optimisation depending on your need. Just do -O0 if you want determinism. But normally you should not rely on it (also at execution time)

Thank you for the summary. My memory of SOTA on disassembly about a year ago was sub—30% accuracy, so this is a significant step forward.

I do think the idea of a 90%+-ish forward and backward assembler LLM is pretty intriguing. There’s bound to be a lot of uses for it; especially if you’re of the mind that to get there it would have to have learned a lot about computers in the foundation model training phase.

Like, you’d definitely want to have those weights somehow baked into a typical coding assistant LLM, and of course you’d be able to automate round one of a lot of historical archiving projects that would like to get compilable modern code but only have a binary, you’d be able to turn some PDP-1 code into something that would compile on a modern machine, … you’d probably be able to leverage it into building chip simulations / code easily, it would be really useful for writing Verilog, (maybe), anyway, the use cases seem pretty broad to me.

Sure, performance is more interesting, but it's significantly harder.

With code size, you just need to run the code through the compiler and you have a deterministic measurement for evaluation.

Performance has no such metric. Benchmarks are expensive and noisy. Cost models seem like a promising direction, but they aren't really there yet.

Maybe they are thinking about embedding a program generator and execution environment into their LLM inferencing loop in a tighter way. The model invents a program that guides the output in a specific/algorithmic way, tailored to the prompt.

some comments like this make me waant to subscribe to you for all your future comments. thanks for doing the hard work of summarizing and taking the bold step of sharing your thoughts in public. i wish more HNers were like you.

You're very kind, thank you. :)

>compilers behaving in a deterministic and predictable way is an important fundamental of pipelines. LLMs are inherently unpredictable, and so using an LLM for compilation / decompilation -- even an LLM that has 99.99% accuracy

You're confusing different concepts here. An llm is technically not unpredictable by itself (at least the ones we are talking about here, there are different problems with beasts like GPT4 [1]). The "randomness" of llms you are probably experiencing stems from the autoregressive completion, which samples from probabilities for a temperature T>0 (which is very common because it makes sense in chat applications). But there is nothing that prevents you from simply choosing greedy sampling, which would make your output 100% deterministic and reproducible. That is particularly useful for disassembling/decompiling and has the chance to vastly improve over existing tools, because it is common knowledge that they are often not the sharpest tools and humans are much better at piecing together working code.

The other question here is accuracy for compiling. For that it is important whether the llm can follow a specification correctly. Because once you write unspecified behaviour, your code is fair game for other compilers as well. So the real question is how well does it follow the spec how good is it at dealing with situations where normal compilers will flounder.

[1] https://152334h.github.io/blog/non-determinism-in-gpt-4/

This was a big unlock for me -- I recall antirez saying the same thing in a comment replying to me when I asked a similar question about potential LLM features in Redis [1].

[1] https://news.ycombinator.com/item?id=39617370

Determinism for any given input isn't an interesting metric though. Inputs are always different, or else you could just replace it with a lookup function. What's important is reliability of the output given a distribution of inputs, and that's where LLMs are unreliable. Temperature sampling can be a technique to improve reliability particularly when things get into repetitve loops - though usually it's to increase creativity.

> What's important is reliability of the output given a distribution of inputs, and that's where LLMs are unreliable.

Nailed it! Thank you for stating this better than I did. Maybe I should have used the word "trustworthy" instead of "predictable".

If I'm using an LLM as a calculator to solve math problems for my users, I may be able to say with certainty that my model is 99% accurate, but I'm not able to know ahead of time which questions it's going to miss, and which ones it's going to get correct. It's a near-infinite input space, so it's difficult to prove correctness via induction. The attention mechanism is (thus far) fairly inscrutable, and thus proving the correctness via deduction is (currently) not possible either.

It would feel weird to use an LLM for a calculator, just as it feels weird to use an LLM as a compiler. I want something that uses an algorithm that is provably and predictably correct -- not one that is "almost always" correct.

That unpredictability of accuracy isn't connected to the temperature used when sampling, so much as the inscrutable nature of the attention mechanism when verifying the trustworthiness over the range of all possible inputs.

> The "randomness" of llms you are probably experiencing stems from the autoregressive completion, which samples from probabilities for a temperature T>0 (which is very common because it makes sense in chat applications).

Even that “random” sampling is deterministic, in that if you use the same PRNG algorithm with the same random seed, then (all else being equal) you should get the same results every time.

To get genuine nondeterminism, you need an external source of randomness, such as thermal noise, keystroke timing, etc. (Even whether that is really non-deterministic depends on a whole lot of contested issues in philosophy and physics, but at least we can say it is non-deterministic for all practical purposes.)

LLMs are probably great at this. You can break down the code into functions or basic blocks. You can use the LLM to decompile them and then test if the decompilation results match the code when executed. You'll probably get this right after a few tries. Then you can train your model with the succesful decompilation results so your model will get better.