// 
// sub_1026F0 from 0x1026f0 to 0x102864 (372 bytes)
// 
// 1 xrefs:
// > <???> @ 0x11a118
// 
// This function is hookable on this platform!
// 
// Using BromaIDA 8.0.0 @ https://github.com/Stazzical/BromaIDA
// Using bindings at commit ba9f177b at Thu Jul 16 22:52:06 2026 from https://github.com/geode-sdk/bindings
// 

float32x4_t *__fastcall sub_1026F0(float32x4_t *result, __int64 a2, __int64 a3, __int64 a4, _QWORD *a5)
{
  __int64 i; // x30
  int v6; // w8
  int8x16_t v7; // q0
  int64x2_t v8; // q1
  int8x16_t v9; // q2
  float32x4_t v10; // q3
  __int64 v11; // x5
  unsigned __int64 v12; // d4
  float32x4_t v13; // q6
  unsigned __int64 v14; // d7
  float v15; // w9
  __int64 v16; // x1
  int v17; // w8

  for ( ; (_DWORD)a2; *(_QWORD *)a4 += *a5 )
  {
    if ( ((unsigned __int8)result & 0xF) == 0 )
      break;
    a2 = (unsigned int)(a2 - 1);
    *result++ = vmlaq_n_f32(
                  *(float32x4_t *)(a3 + 4LL * (unsigned int)(4 * *(_DWORD *)(a4 + 4))),
                  vsubq_f32(
                    *(float32x4_t *)(a3 + 4LL * (unsigned int)(4 * *(_DWORD *)(a4 + 4) + 4)),
                    *(float32x4_t *)(a3 + 4LL * (unsigned int)(4 * *(_DWORD *)(a4 + 4)))),
                  (float)(*(_DWORD *)a4 >> 1) * 4.6566e-10);
  }
  if ( (int)a2 >> 2 )
  {
    result = (float32x4_t *)sub_102CE0(result, a2, a3, a4, a5, i);
    for ( i = v11; v6; result += 4 )
    {
      v12 = vshl_n_s32(vzip2_s32((int32x2_t)v7.n128_u64[0], (int32x2_t)vextq_s8(v7, v7, 8u).n128_u64[0]), 2u).n64_u64[0];
      v13 = vmulq_f32(vcvtq_f32_u32(vshrq_n_u32(vuzp1q_s32(v7, v9), 1u)), v10);
      *result = vmlaq_n_f32(
                  *(float32x4_t *)(a3 + 4LL * (unsigned int)v12),
                  vsubq_f32(
                    *(float32x4_t *)(a3 + 4LL * (unsigned int)v12 + 16),
                    *(float32x4_t *)(a3 + 4LL * (unsigned int)v12)),
                  v13.n128_f32[0]);
      v14 = vshl_n_s32(vzip2_s32((int32x2_t)v9.n128_u64[0], (int32x2_t)vextq_s8(v9, v9, 8u).n128_u64[0]), 2u).n64_u64[0];
      result[1] = vmlaq_lane_f32(
                    *(float32x4_t *)(a3 + 4LL * HIDWORD(v12)),
                    vsubq_f32(*(float32x4_t *)(a3 + 4LL * HIDWORD(v12) + 16), *(float32x4_t *)(a3 + 4LL * HIDWORD(v12))),
                    (float32x2_t)v13.n128_u64[0],
                    1);
      --v6;
      result[2] = vmlaq_laneq_f32(
                    *(float32x4_t *)(a3 + 4LL * (unsigned int)v14),
                    vsubq_f32(
                      *(float32x4_t *)(a3 + 4LL * (unsigned int)v14 + 16),
                      *(float32x4_t *)(a3 + 4LL * (unsigned int)v14)),
                    v13,
                    2);
      v7 = vaddq_s64(v7, v8);
      v9 = vaddq_s64(v9, v8);
      result[3] = vmlaq_laneq_f32(
                    *(float32x4_t *)(a3 + 4LL * HIDWORD(v14)),
                    vsubq_f32(*(float32x4_t *)(a3 + 4LL * HIDWORD(v14) + 16), *(float32x4_t *)(a3 + 4LL * HIDWORD(v14))),
                    v13,
                    3);
    }
    *(_QWORD *)a4 = v7.n128_u64[0];
  }
  v15 = 4.6566e-10;
  if ( (a2 & 3) != 0 )
  {
    do
    {
      *result = vmlaq_n_f32(
                  *(float32x4_t *)(a3 + 4LL * (unsigned int)(4 * *(_DWORD *)(a4 + 4))),
                  vsubq_f32(
                    *(float32x4_t *)(a3 + 4LL * (unsigned int)(4 * *(_DWORD *)(a4 + 4) + 4)),
                    *(float32x4_t *)(a3 + 4LL * (unsigned int)(4 * *(_DWORD *)(a4 + 4)))),
                  (float)(*(_DWORD *)a4 >> 1) * v15);
      result = (float32x4_t *)sub_102E10(&result[1], i);
      i = v16;
    }
    while ( v17 );
  }
  return result;
}