Interesting. 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.
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.
>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.
Having a way to assert that the compiler does what you expect can be helpful in large projects with many contributors of different skill level. Having something fail when a random change breaks autovectorization can save a lot of time profiling. When a compiler upgrade changes codegen I would also prefer an assertion telling me about it so that I can run relevant benchmarks to see whether it’s an improvement or not. Relying on whole-system benchmarks is difficult due to noise.
It feels like a leaky abstraction, and similarly implies there should be some way to drop down a layer and work directly with what’s beneath. Something in between C and non-portable assembly.
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?
I have written a compiler for a language at more or less this abstraction level. It provides access to 16 general purpose registers and a set of virtual instructions that operate on these registers. I program on an intel mac so the virtual instructions all map directly to x86_64 instructions, but it would be very straightforward to write an arm backend that may compose multiple arm instructions for the equivalent x86_64 behavior. I also could support virtual registers for platforms with fewer than 16 registers.
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.
I think a better approach might be automated performance regression tests. That's checking the property you probably actually care about directly (performance) and leaves the compiler (and other engineers) some leeway to do better without breaking the test.
Actually setting up a robust system for perf regression tests is tricky though...
It is brittle and a leaky abstraction. Code is already too complex, and correctness and time complexity are far more important than these micro-optimizations. Although I think your language feature seems reasonable.
It depends on your use case, as always. Correctness is not always black and white (hence my favourite compilation flag, -funsafe-math-optimizations) and time complexity can be misleading (O(log N) with a large base is O(1) in practice), but a correct, theoretically optimal algorithm might still be leaving a lot of performance on the table. If you're a hyperscaler, a high-frequency trader, or perhaps a game programmer pushing the limits of the platform, small gains can accumulate meaningfully, as can the thousand cuts caused by small inefficiencies spread out over the codebase.
> our optimisation relies on having r < count, but if count is zero this condition cannot be met. [Since taking the remainder of a division by zero is undefined behaviour, the compiler would be well within its rights to assume that count != 0 and proceed with the transformation.]
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.
BTW. On ARM and RISC-V, integer remainder from division by zero does indeed result in the dividend (and no trap), unless the compiler is doing something unusual.
> Here, the simplification analysis uses the conditional hidden in the assert() macro to figure out that we only execute the urem instruction if cur < count (see the isImpliedByDomCondition() method). This can have the unexpected consequence of making builds with assertions enabled faster in some cases; we can restore performance by replacing disabled assertions with assume attributes, which have no runtime impact.
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.
> 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.
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".
Yeah, good point. Honestly it's been so long since I've added that to a project (it's normally hidden in some include that everything else includes) that I'd forgotten it wasn't a compiler level reserved keyword for C++ code.
> 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.
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.
Interesting. 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.
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.
(Author here)
>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.
Having a way to assert that the compiler does what you expect can be helpful in large projects with many contributors of different skill level. Having something fail when a random change breaks autovectorization can save a lot of time profiling. When a compiler upgrade changes codegen I would also prefer an assertion telling me about it so that I can run relevant benchmarks to see whether it’s an improvement or not. Relying on whole-system benchmarks is difficult due to noise.
Haskell has a package to make testing this sort of thing easier https://hackage.haskell.org/package/inspection-testing
It feels like a leaky abstraction, and similarly implies there should be some way to drop down a layer and work directly with what’s beneath. Something in between C and non-portable assembly.
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?
I have written a compiler for a language at more or less this abstraction level. It provides access to 16 general purpose registers and a set of virtual instructions that operate on these registers. I program on an intel mac so the virtual instructions all map directly to x86_64 instructions, but it would be very straightforward to write an arm backend that may compose multiple arm instructions for the equivalent x86_64 behavior. I also could support virtual registers for platforms with fewer than 16 registers.
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.
> Something in between C and non-portable assembly.
Writing your own LLVM passes?
I think a better approach might be automated performance regression tests. That's checking the property you probably actually care about directly (performance) and leaves the compiler (and other engineers) some leeway to do better without breaking the test.
Actually setting up a robust system for perf regression tests is tricky though...
It is brittle and a leaky abstraction. Code is already too complex, and correctness and time complexity are far more important than these micro-optimizations. Although I think your language feature seems reasonable.
It depends on your use case, as always. Correctness is not always black and white (hence my favourite compilation flag, -funsafe-math-optimizations) and time complexity can be misleading (O(log N) with a large base is O(1) in practice), but a correct, theoretically optimal algorithm might still be leaving a lot of performance on the table. If you're a hyperscaler, a high-frequency trader, or perhaps a game programmer pushing the limits of the platform, small gains can accumulate meaningfully, as can the thousand cuts caused by small inefficiencies spread out over the codebase.
> our optimisation relies on having r < count, but if count is zero this condition cannot be met. [Since taking the remainder of a division by zero is undefined behaviour, the compiler would be well within its rights to assume that count != 0 and proceed with the transformation.]
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.
BTW. On ARM and RISC-V, integer remainder from division by zero does indeed result in the dividend (and no trap), unless the compiler is doing something unusual.
> r + 1 == count ? 0 : r + 1
It'is legal!
But looks like a few more steps when compiled.
> Here, the simplification analysis uses the conditional hidden in the assert() macro to figure out that we only execute the urem instruction if cur < count (see the isImpliedByDomCondition() method). This can have the unexpected consequence of making builds with assertions enabled faster in some cases; we can restore performance by replacing disabled assertions with assume attributes, which have no runtime impact.
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.
> 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.
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".
Yeah, good point. Honestly it's been so long since I've added that to a project (it's normally hidden in some include that everything else includes) that I'd forgotten it wasn't a compiler level reserved keyword for C++ code.
> 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.
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.
My point is that even though the assert() is optimised out, the compiler could still assume that the condition is valid.
How? The assert() has no special significance to the compiler at all, it's just code. Usually it's just an empty macro for non-debug builds.
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).
There's an interesting comment on this over on Lobsters: https://lobste.rs/s/e4y5ps/two_studies_compiler_optimisation...
Do not check OP's art on a work machine lol
[dead]
[dead]
[flagged]
[flagged]