When compilers surprise you
xania.orgThe code that does this is here, if anyone is curious:
https://github.com/llvm/llvm-project/blob/release/21.x/llvm/...
https://github.com/llvm/llvm-project/blob/release/21.x/llvm/...
Almost 16000 lines in a single source code file. I find this both admirable and unsettling.
Does it really matter where the lines are? 16,000 lines is still 16,000 lines.
Even though I do find your indifference refreshing I must say: it does matter for quite a few people.
If you want recognize all the common patterns, the code can get very verbose. But it's all still just one analysis or transformation, so it would be artificial to split into multiple files. I haven't worked much in llvm, but I'd guess that the external interface to these packages is pretty reasonable and hides a large amount of the complexity that took 16kloc to implement
If you don’t rely on IDE features or completion plugins in an editor like vim, it can be easier to navigate tightly coupled complexity if it is all in one file. You can’t really scan it or jump to the right spot as easily as smaller files, but in vim searching for the exact symbol under the cursor is a single character shortcut, and that only works if the symbol is in the current buffer. This type of development works best for academic style code with a small number (usually one or two) experts that are familiar with the implementation, but in that context it’s remarkably effective. Not great for merge conflicts in frequently updated code though.
* requires two key presses which is identical to <C-]> or g]> but in vim searching for the exact symbol under the cursor is a single character shortcuthttps://kulkarniamit.github.io/whatwhyhow/howto/use-vim-ctag...
... yes.
If it was 16K lines of modular "compositional" code, or a DSL that compiles in some provably-correct way, that would make me confident. A single file with 16K lines of -- let's be honest -- unsafe procedural spaghetti makes me much less confident.
Compiler code tends to work "surprisingly well" because it's beaten to death by millions of developers throwing random stuff at it, so bugs tend to be ironed out relatively quickly, unless you go off the beaten path... then it rapidly turns out to be a mess of spiky brambles.
The Rust development team for example found a series of LLVM optimiser bugs related to (no)aliasing, because C/C++ didn't use that attribute much, but Rust can aggressively utilise it.
I would be much more impressed by 16K lines of provably correct transformations with associated Lean proofs (or something), and/or something based on EGG: https://egraphs-good.github.io/
On the other end of the optimizer size spectrum, a surprising place to find a DSL is LuaJIT’s “FOLD” stage: https://github.com/LuaJIT/LuaJIT/blob/v2.1/src/lj_opt_fold.c (it’s just pattern matching, more or less, that the DSL compiler distills down to a perfect hash).
Part of the issue is that it suggests that the code had a spaghettified growth; it is neither sufficient nor necessary but lacking external constraints (like an entire library developed as a single c header) it suggests that code organisation is not great.
Hardware is often spaghetti anyway. There are a large number of considerations and conditions that can invalidate the ability to use certain ops, which would change the compilation strategy.
The idea of good abstractions and such falls apart the moment the target environment itself is not a good abstraction.
I was not expressing an opinion on giant files, I was just postulating on why people dislike them.
I find the real question: are all 16,000 of those lines require to implement the optimization? How much of that is dealing with LLVM’s internal representation and the varying complexity of LLVM’s other internal structure?
What would it mean to implement the optimization without dealing with LLVM's internal structure? Optimizations don't exist in a vacuum.
I do too, but I'm pretty sure I've seen worse.
Thank you, bumholes
These sorts of things are fun and interesting. Compiler optimizations fall into two categories:
1. organized data flow analysis
2. recognizing a pattern and replacing it with a faster version
The first is very effective over a wide range of programs and styles, and is the bulk of the actual transformations. The second is a never-ending accumulation of patterns, where one reaches diminishing returns fairly quickly.
The example in the linked article is very clever and fun, but not really of much value (I've never written a loop like that in 45 years). As mentioned elsewhere "Everyone knows the Gauss Summation formula for sum of n integers i.e. n(n+1)/2" and since everyone knows it why not just write that instead of the loop!
Of course one could say that for any pattern, like replacing i*2 with i<<1, but those pattern replacements are very valuable because they are generated by high level generic coding.
And you could say I'm just being grumpy about this because my optimizer does not do this particular optimization. Fair enough!
It might have more value than you think. If you look up SCEV in LLVM you'll see it's primarily used for analysis and it enables other optimizations outside of math loops that, by themselves, probably don't show up very often.
You might be right.
It's not clear to me what optimizations the compiler actually did here. Years ago, I worked on a niche compiler, and was routinely surprised by what the optimizer was able to figure out; despite having personally written most of the optimization transformations myself.
I can't actually speak to the specifics here but usually this is "idiom recognition", that is, it just notices that the pattern is there and transforms it directly.
The point is not really that the compiler is able to spot the Gaussian sum. It uses a general infrastructure to turn many "math loops" that sum/multiply numbers into closed-form solutions. This handles Gauss but is much more general. In turn it allows you to write simple code that still does the right thing, instead of a complicated-looking large multiplication with a bunch of random-seeming constants.
isn't that kind of stuff mostly useful for generated code?
One recognizes different patterns for both the intermediate representation and the generated code.
That one is called scalar evolution, llvm abbreviates it as SCEV. The implementation is relatively complicated.
More similar optimizations: https://matklad.github.io/2025/12/09/do-not-optimize-away.ht...
The beginning of that article is slightly wrong: the compiler should compute N(N-1)/2 (and does), because the original code adds up all the numbers from 0 to N excluding N. The usual formulation in math includes the upper bound: the sum of integers from 1 to N, including N, is N(N+1)/2, so you have to replace N by (N-1) if you want a formula for the sum where the last number is N-1.
Couldn't the compiler optimise this still? Make two versions of the function, one with constant folding and one without. Then at runtime, check the value of the parameter and call the corresponding version.
Yes, a sufficiently smart compiler can always tell you’re doing a benchmark and delete it. It’s just unlikely.
What's actually way cooler about this is that it's generic. Anybody could pattern match the "sum of a finite integer sequence" but the fact that it's general purpose is really awesome.
Compilers can add way more closed forms. Would it be worth it?
I'm once again surprised at GCC being slower than clang. I would have thought that GCC, which had a 20? year head start would've made faster code. And yet, occasionally I look into the assembly and go "what are you doing?" And the same flags + source into clang is better optimized or uses better instructions or whatever. One time it was bit extraction using shifts. Clang did it in 2 steps: shift left, shift right. GCC did it in 3 I think? I think it maybe shifted right first or maybe did a logical instead of arithmetic and then sign extended. Point is, it was just slower.
Compiler know-how and resources available during compilations made very signicant progress between gcc and LLVM/clang era.
gcc was and is an incredible achievement, but it is traditionally considered difficult to implement many modern compiler techqniques in it. It's at least unpleasant, let's put it this way.
Not sure whether this is generally true. GCC appears to have similar optimizations and I personally find LLVM's code much more intimidating. But it is certainly true that LLVM seems to see more investment. I assume the license may also play a role. For comparison, here is some related code:
https://github.com/gcc-mirror/gcc/blob/master/gcc/tree-chrec... https://github.com/llvm/llvm-project/blob/release/21.x/llvm/...
GCC has almost the same modern compiler techniques implemented.
GCC and Clang are largely similar when it comes to performance as each implements passes the other does not. It’s always possible to find examples where they optimize a piece of code differently and one comes out ahead of the other.
Did it involve bitfields? GCC is notoriously bad at optimizing them. There are some target-specific optimizations, but pretty much nothing in the middle-end.
It did, yes. On an architecture without bit field extracts.
I'm not. GCC started out as a work of idealistic licensing purists and was deliberately "obfuscated" to make it hard to extend and embed. That stance has since been softened considerably, but the code generator is still far more complex than it needs to be, and I think that has made it harder to modify for efficiency. Clang is far less ideology-focused and its structure makes implementing optimisations easier.
On the other hand, I find MSVC and especially ICC output to be quite decent, although I have never seen their source code.
Having inspected the output of compilers for several decades, it's rather easy to tell them apart.
It’s neat. I wonder if someone attempted detecting a graph coloring problem to replace it with a constant.
Graph coloring is NP-hard so it would be very difficult to replace it with an O(1) algorithm.
If you mean graph coloring restricted to planar graphs, yes it can always be done with at most 4 colors. But it could still be less, so the answer is not always the same.
(I know it was probably not a very serious comment but I just wanted to infodump about graph theory.)
This is really bluring the line between implementation and specification. You may think you're writing the implementation but it is really a proxy for the specification. In other words, the compiler creating an illusion of an imperative machine.
I will admit I was initially surprised Matt was not already familiar with this behavior given his reputation. I remember discovering it while playing with llvm intermediate representation 10 years ago in college. I would never have considered myself very knowledgeable about modern compilers, and have never done any serious performance work. In that case it had solved a recursion to a simple multiplication, which completely surprised me. The fact that Matt did not know this makes me think this pass may only work on relatively trivial problems that he would never have written in the first place, and therefore never have witnessed the optimization.
He was: he brought up the very same example in a talk in 2017.
Ah that makes much more sense. I guess he means the optimization is surprising when you first discover it, which it certainly was for me!
That's neat.
A hard problem in optimization today is trying to fit code into the things complex SSE-type instructions can do. Someone recently posted an example where they'd coded a loop to count the number of one bits in a word, and the compiler generated a "popcount" instruction. That's impressive.
It may be a different post, but I covered this earlier this month in the same series of blog posts/YouTube videos.
I'm actually surprised that gcc doesn't do this! If there's one thing compilers do well is pattern match on code patterns and replace with more efficient ones; just try pasting things from Hacker's Delight and watch it always canonicalise it to the equivalent, fastest machine code.
This particular case isn't really due to pattern matching -- it's a result of a generic optimization that evaluates the exit value of an add recurrence using binomial coefficients (even if the recurrence is non-affine). This means it will work even if the contents of the loop get more exotic (e.g. if you perform the sum over x * x * x * x * x instead of x).
Doing something like that with a pattern is obvious, but also useless, as it will catch very limited cases. The example presented, is known there is a closed form (it’s believed Gauss even discovered it being 6 yo). I’m sure this optimization will catch many other things, so is not trivial at all.
First time I encountered that book was seeing it on the desk of a compiler engineer.
A lot of hardcoding, making expression consistent, e.g transforming a+3 into 3+a for easier pattern matching
If you now have a function where you call this one with an integer literal, you will end up with a fully inlined integer answer!
Could do that whether SCEV’d or not with C++20 consteval, lol.
The first thing I had in mind was: the final answer needed to be /2. keeping the number before dividing not overflowing needs some tedious work
It's not very tedious. Instead of dividing the product by 2, you can just divide whichever of x or x+1 is even by 2 before multiplying.
Only thing that surprised me was that GCC didn't manage to optimize it. I expected it to be able to do so.
> I love that despite working with compilers for more than twenty years, they can still surprise and delight me.
This kind of optimization, complete loop removal and computing the final value for simple math loops, is at least 10 years old.
10 years is not a lot. Is almost “yesterday” things being done in a field 10 years old, can still surprise experts in the field. With 30+ years experience I still find relatively new things, that are maybe 15 yo.
In topics like compiler optimization, is not like there are many books which describe this kind of algorithms.
Learning something old can be surprising. Enjoying that learning can be delightful.
Seems like the author is both surprised and delighted with an optimization they learned of today. Surely you’ve been in the same situation before.
This exact content was posted a few months ago. Is this AI or just a copy paste job?
You're probably thinking of another post (https://xania.org/202512/11-pop-goes-the-weasel-er-count) where an entire loop was optimized to a single instruction
This exact content was only posted today? :)