Rust compiler or LLVM fails to generate branch-free code

[Sword-Smith]

Not a bug but a missed opportunity to generate fast code.

In some instances the Rust compiler generates faster code when if/else branching is used than when "branch-free" arithmetical functions are used.

This results in code that's significantly slower than it should be.

This problem is documented in the following PR.
Neptune-Crypto/twenty-first#70

[Sword-Smith]

It's hard for me to add a minimal example since in isolation both options generate branch-free code. It's only when it's wrapped in a larger context that the discrepancy arises

Here's an example of code where branch-free code generation happens in both cases: https://rust.godbolt.org/z/KzedKxfW5

An example where the problem arises can be found by cloning the above mentioned repository.

[workingjubilee]

This is almost certainly caused by an LLVM pass that deliberately introduces branches. In some cases, this is an optimization. I recommend reporting this issue upstream.