Bitcast

A bitcast preserves bits; a numeric conversion preserves a numeric value as far as the destination can represent it. They are not interchangeable. For example, bitcasting 1.0f32 to i32 yields its IEEE bit encoding, not integer one.

This reference exercise checks the distinction within a complete feature context. Integer/float bit manipulation is used by zen color and approximation kernels; this roundtrip is an API test rather than a production algorithm.

use archmage::prelude::*;
#[magetypes(define(f32x8), v3, neon, wasm128, scalar)]
fn bits_impl(token: Token, input: [f32; 8]) -> ([i32; 8], [f32; 8]) {
    let v = f32x8::from_array_t(token, input);
    let bits = v.bitcast_to_i32();
    (bits.to_array(), bits.bitcast_to_f32().to_array())
}
pub fn bits(input: [f32; 8]) -> ([i32; 8], [f32; 8]) {
    incant!(bits_impl(input), [v3, neon, wasm128, scalar])
}
let (encoded, roundtrip) = bits([1.0; 8]);
assert_eq!(encoded, [1.0f32.to_bits() as i32; 8]);
assert_eq!(roundtrip, [1.0; 8]);

Use bitcast_to_i32 / bitcast_to_f32 for value casts. Older width-specific names such as bitcast_i32x8 remain compatibility aliases. Signed/unsigned byte casts have shape-specific names; follow their method reference.

Reference casts preserve borrowing and require compatible size/alignment and valid bit patterns; use only the library's provided methods. They do not make arbitrary user structs safe to reinterpret. Prefer value operations unless a measured caller needs a borrowed view.