use crate::fast_divide::{DividerI32, rounded_quotients_scalar};
use core::arch::wasm32::*;
#[inline]
#[target_feature(enable = "simd128")]
fn mulhi_i32x4(m: v128, magic: v128) -> v128 {
let lo = i64x2_extmul_low_i32x4(m, magic);
let hi = i64x2_extmul_high_i32x4(m, magic);
i32x4_shuffle::<1, 3, 5, 7>(lo, hi)
}
#[inline]
#[target_feature(enable = "simd128")]
fn magic_quotients<const ADD: bool>(v4s: &mut [[i32; 4]], half: i32, divider: DividerI32) {
let half = i32x4_splat(half);
let magic = i32x4_splat(divider.magic());
let shift = divider.shift();
for v4 in v4s {
let v = unsafe { v128_load(v4.as_ptr().cast()) };
let sign = i32x4_shr(v, 31);
let m = i32x4_add(i32x4_abs(v), half);
let mut uq = mulhi_i32x4(m, magic);
if ADD {
uq = i32x4_add(uq, m);
}
let q = i32x4_shr(uq, shift);
let q = i32x4_sub(q, i32x4_shr(q, 31));
let out = i32x4_sub(v128_xor(q, sign), sign);
unsafe { v128_store(v4.as_mut_ptr().cast(), out) };
}
}
#[target_feature(enable = "simd128")]
pub(crate) fn rounded_quotients_wasm(plane: &mut [i32], q: i32) {
debug_assert!(q > 1);
let divider = DividerI32::new(q);
let (v4s, tail) = plane.as_chunks_mut::<4>();
if divider.magic() == 0 {
let half = i32x4_splat(q / 2);
let shift = divider.shift();
for v4 in v4s.iter_mut() {
let v = unsafe { v128_load(v4.as_ptr().cast()) };
let sign = i32x4_shr(v, 31);
let r = i32x4_shr(i32x4_add(i32x4_abs(v), half), shift);
let out = i32x4_sub(v128_xor(r, sign), sign);
unsafe { v128_store(v4.as_mut_ptr().cast(), out) };
}
} else if divider.adds_numerator() {
magic_quotients::<true>(v4s, q / 2, divider);
} else {
magic_quotients::<false>(v4s, q / 2, divider);
}
rounded_quotients_scalar(tail, q);
}