rssn-advanced 0.1.0

//! `cmp_eq_f64x4` — packed `f64x4` equality test.
//!
//! Returns a 4-lane boolean mask: `0xFF` per lane if equal, `0x00`
//! otherwise. `NaN == NaN` is false (IEEE 754), matching Rust `==`.
//!
//! * `x86_64` + AVX2: `vcmpeqpd ymm` + `vmovmskpd` to collapse sign bits.
//! * `AArch64`: two `fcmeq v.2d` + `umov` to extract lane masks (NEON mandatory).
//! * riscv64 + RVV: `vmfeq.vv` to produce a mask register, then `vmerge.vim`
//!   with e8/vl=4 to expand each mask bit to `0xFF` or `0x00`.
//! * fallback: scalar `==` loop.

#![allow(unsafe_code)]

/// Compares `lhs` and `rhs` element-wise, writing `0xFF`/`0x00` into
/// `mask` per lane.
///
/// # Safety
///
/// AVX2 path uses raw pointers and 256-bit unaligned ops, plus a
/// `vmovmskpd` to extract the per-lane sign-bit mask. Lengths are
/// checked; alignment isn't.
#[allow(clippy::inline_always)]
#[inline(always)]
pub fn apply(lhs: &[f64], rhs: &[f64], mask: &mut [u8]) {
    debug_assert!(
        lhs.len() == 4 && rhs.len() == 4 && mask.len() == 4,
        "cmp_eq_f64x4::apply requires exactly 4-element slices \
         (got lhs={}, rhs={}, mask={})",
        lhs.len(),
        rhs.len(),
        mask.len()
    );
    if lhs.len() != 4 || rhs.len() != 4 || mask.len() != 4 {
        return;
    }

    #[cfg(target_arch = "x86_64")]
    {
        if std::is_x86_feature_detected!("avx2") {
            // SAFETY: lengths checked; AVX2 detected.
            unsafe {
                use core::arch::asm;
                // After `vcmpeqpd`, each ymm0 lane is all-ones (0xFFF…)
                // for equal, all-zeros for unequal. We then collapse it
                // to a 4-bit mask via `vmovmskpd`, and explode that back
                // into 4 bytes by hand for the caller.
                let mut packed: u32 = 0;
                asm!(
                    "vmovupd ymm0, ymmword ptr [{lhs}]",
                    "vmovupd ymm1, ymmword ptr [{rhs}]",
                    "vcmpeqpd ymm0, ymm0, ymm1",
                    "vmovmskpd {packed:e}, ymm0",
                    lhs = in(reg) lhs.as_ptr(),
                    rhs = in(reg) rhs.as_ptr(),
                    packed = out(reg) packed,
                    out("ymm0") _,
                    out("ymm1") _,
                    options(nostack, readonly),
                );
                for (i, slot) in mask.iter_mut().enumerate() {
                    *slot = if packed & (1 << i) != 0 { 0xFF } else { 0x00 };
                }
            }
            return;
        }
    }

    #[cfg(target_arch = "aarch64")]
    {
        // SAFETY: lengths checked above; NEON is mandatory on AArch64.
        // `fcmeq` sets each 64-bit lane to all-ones (!=0) if equal, all-zeros (0) otherwise.
        let m0: u64;
        let m1: u64;
        let m2: u64;
        let m3: u64;
        unsafe {
            use core::arch::asm;
            asm!(
                "ld1 {{v0.2d}}, [{lhs}]",
                "ld1 {{v1.2d}}, [{rhs}]",
                "fcmeq v0.2d, v0.2d, v1.2d",
                "umov {m0}, v0.d[0]",
                "umov {m1}, v0.d[1]",
                "ld1 {{v0.2d}}, [{lhs}, #16]",
                "ld1 {{v1.2d}}, [{rhs}, #16]",
                "fcmeq v0.2d, v0.2d, v1.2d",
                "umov {m2}, v0.d[0]",
                "umov {m3}, v0.d[1]",
                lhs = in(reg) lhs.as_ptr(),
                rhs = in(reg) rhs.as_ptr(),
                m0 = out(reg) m0,
                m1 = out(reg) m1,
                m2 = out(reg) m2,
                m3 = out(reg) m3,
                out("v0") _,
                out("v1") _,
                options(nostack, readonly),
            );
        }
        let lane_bits = [m0, m1, m2, m3];
        for (slot, &bits) in mask.iter_mut().zip(lane_bits.iter()) {
            *slot = if bits != 0 { 0xFF } else { 0x00 };
        }
        return;
    }

    #[cfg(all(target_arch = "riscv64", target_feature = "v"))]
    {
        // RVV 1.0 with vl=4, e64 for the comparison, then vl=4, e8 to
        // expand each mask bit to 0xFF / 0x00 in one `vmerge.vim`.
        //
        // `vmfeq.vv v0, v1, v2` writes a *mask register* (v0.m) where
        // bit i = 1 iff lhs[i] == rhs[i] (NaN != NaN per IEEE-754).
        // Switching SEW to e8 with the same vl=4 keeps the mask bits
        // aligned: bit i of v0 still corresponds to element i.
        // `vmerge.vim v1, v1, -1, v0` writes 0xFF (simm5=-1 sign-extends
        // to all-ones for e8) where the mask bit is 1, 0x00 otherwise.
        unsafe {
            use core::arch::asm;
            asm!(
                // Phase 1: compare as f64x4, result in mask register v0.
                "li t0, 4",
                "vsetvli t0, t0, e64, m1, ta, ma",
                "vle64.v v1, ({lhs})",
                "vle64.v v2, ({rhs})",
                "vmfeq.vv v0, v1, v2",
                // Phase 2: expand mask bits to 0xFF/0x00 bytes.
                "vsetvli t0, t0, e8, m1, ta, ma",
                "vmv.v.i v1, 0",
                "vmerge.vim v1, v1, -1, v0",
                "vse8.v v1, ({mask})",
                lhs  = in(reg) lhs.as_ptr(),
                rhs  = in(reg) rhs.as_ptr(),
                mask = in(reg) mask.as_mut_ptr(),
                out("t0") _,
                out("v0") _,
                out("v1") _,
                out("v2") _,
                options(nostack),
            );
        }
        return;
    }

    // Bitwise equality on f64 is intentional: callers want IEEE-754
    // `==` semantics where NaN != NaN.
    #[allow(clippy::float_cmp)]
    for ((l, r), slot) in lhs.iter().zip(rhs.iter()).zip(mask.iter_mut()) {
        *slot = if *l == *r { 0xFF } else { 0x00 };
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn equal_lanes_set_to_ff() {
        let a = [1.0_f64, 2.0, 3.0, 4.0];
        let b = [1.0_f64, 0.0, 3.0, 4.0];
        let mut m = [0_u8; 4];
        apply(&a, &b, &mut m);
        assert_eq!(m, [0xFF, 0x00, 0xFF, 0xFF]);
    }

    #[test]
    fn nan_compares_not_equal() {
        let nan = f64::NAN;
        let a = [nan, 0.0, nan, 1.0];
        let b = [nan, 0.0, 1.0, nan];
        let mut m = [0_u8; 4];
        apply(&a, &b, &mut m);
        assert_eq!(m[0], 0x00, "NaN == NaN must be false (IEEE 754)");
        assert_eq!(m[1], 0xFF);
        assert_eq!(m[2], 0x00);
        assert_eq!(m[3], 0x00);
    }
}