Two studies in compiler optimisations
hmpcabral.com
hmpcabral.com
Not picking on the OA in the slightest; just thinking in terms of holistic system design. If you know what you want to happen, and you are smart enough to introspect the behavior of the tool and decide that it didnt happen, you are more than smart enough to just write it correctly in the first place.
Perhaps that is unrealistic, perhaps there is a hidden iceberg of necessary but convolutive optimizations no human could realistically or legibly write. But ok, where do you really need to engage in this kind of optimization golf? Inlined functions?
Ok, what about this targeted language feature for a future-day Zig:
1. Write an ordinary zig function 2. Write inline assembly version of that function 3. Write a "comptime assert" that first compiles to second, which only "runs" for the relevant arch. 4. What should that assert mean? That the compiler just uses your assembly version instead, but _also_ uses existing compiler machinery or an external theorem prover to verify they "behave the same up to X", for customizable values of X
That has the right feel, maybe. You are "pinning" specific, vetted, optimizations without compromising the intent, readability, or correctness of your code. And easy iteration is possible, because a failing comptime assert will just dump the assembly; you can even start with an empty manual impl.
Asserting retroactively that compilers produce the correct assemvbly feels like just plain giving up on everything in between. Surely the best we can do isn’t a bunch of flaky weirdly interacting optimizations, UB footguns everywhere, things changing when updating the toolchain, etc?
Writing your own LLVM passes?
Having this compiler, which is extremely fast (it is self hosted and self compiles its own 9750K line source file in 15ms on my 2019 macbook pro and I've seen throughout ranging from 500K to 10M lines per second for other programs), I have little interest in ever using LLVM again. I would rather just write optimal code than pray that LLVM does a good job for me. For equivalent programs, I find LLVM can be anywhere from 50-10000x slower than my compiler and doesn't necessarily produce better code. I can produce examples of LLVM code becoming significantly worse at -O2 than at -O0.
The only real utility LLVM offers me at this point is that I could use it to figure out what an optimal solution to a small subprogram might be by compiling a small c program and inspecting the output. Then I can just directly write the optimal solution instead of being perpetually bogged down by llvm's bloat as well as also being exposed to unexpected regressions in the optimizer and/or having to learn its strange incantations to get desired performance.
Actually setting up a robust system for perf regression tests is tricky though...
>It all seems very brittle, though. And that something has gone very wrong with our ecosystem of tools, languages, and processes when it becomes advisable to massage source until specific passes in a specific version LLVM don't mess things up for other passes.
I would say the main takeaway is actually to not do that, precisely because it's brittle, difficult to understand, and can backfire by making things harder for the compiler. As I point out at the end, the vast majority of developers will be better served by sticking to common idioms and making their intent as clear as possible using language facilities, rather than trying to outsmart the compiler or, conversely, relying excessively on the optimiser as a magic black box.
That being said, I do find it helpful to understand the broad strokes of the optimisation pipeline (things like "callees are optimised before callers", "maintaining invariants enables more simplifications", or "early simplifications are better") to make the most of it. Like with any other tool, mastery means letting it do its job most of the time but knowing when and where to step in if necessary.
Interestingly, even if you define division by zero to produce the remainder equal to the dividend, this optimization (replacing "(r + 1) % count" with "r + 1 == count ? 0 : r + 1") would still be legal.
It'is legal!
But looks like a few more steps when compiled.
It surprises me that the compiler doesn't still take the inference from the assert and just disable emitting the code to perform the check. Over the last 15 years I've worked on many codebases that are written with unnecessary asserts, partly as documentation, but maybe because people assumed it helped the compiler.
I've also worked on many codebases (and written code like this on my own projects) where the code looks like: assert(condition); if (condition) { ... } because I still want that safety check in release builds, and want the exception in debug builds but absolutely not ever in release builds.
The compiler isn't as clever as I think you're envisioning: assert() only works that way because it exits the control flow if the statement isn't true.
I guess you could define your assert() use [[assume]] in C++ for non-debug builds... but that seems like a very bad idea to me.
Just leave the asserts in prod. They almost certainly don't have measurable overhead. The few that do can be dealt with separately.
The Linux kernel has thousands of asserts which are always checked at runtime (BUG_ON).
That's because that's what the <assert.h> assert() must do; it's specified to do and imply nothing when assertions are disabled. (the standard literally fully defines it as `#define assert(...) ((void)0)` when NDEBUG)
Whereas `[[assume(...)]]` is a thing specifically for that "infer things from this without actually emitting any code".