Eliminating Bounds Checks Safely & CPU PMU Profiling

The Cost of Runtime Bounds Checks

To guarantee memory safety, Rust automatically inserts bounds checks (if index >= len { panic!() }) on slice indexing operations (slice[i]). In tight mathematical and tokenization loops, these branch instructions prevent the compiler from auto-vectorizing with SIMD.

Idiomatic Ways to Eliminate Bounds Checks

1. Iterators instead of manual indexing: for item in slice generates zero bounds checks.
2. Iterator Chunking & Windows: slice.chunks_exact(8) proves to the compiler that each chunk has exactly 8 elements.
3. Asserting Slices Upfront: Calling assert!(slice.len() >= 1024); upfront allows LLVM to prove that subsequent indexes within that range cannot fail, eliminating all 1024 individual checks inside the loop!

// Hoisting bounds checks out of inner loops
pub fn process_tensor_slice(data: &mut [f32]) {
    assert!(data.len() >= 4); // Single bounds check here
    
    // LLVM recognizes data.len() >= 4 and strips bounds checks below!
    data[0] *= 2.0;
    data[1] *= 2.0;
    data[2] *= 2.0;
    data[3] *= 2.0;
}