// 
// sub_1010C0 from 0x1010c0 to 0x101a98 (2520 bytes)
// 
// 1 xrefs:
// > sub_101080 @ 0x1010a4
// 
// 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
// 

int8x16_t *__usercall sub_1010C0@<X0>(
        int8x16_t *result@<X0>,
        int8x16_t *a2@<X1>,
        int a3@<W2>,
        float32x4_t *a4@<X3>,
        int a5@<W4>,
        int a6@<W5>,
        int a7@<W8>,
        int8x16_t a8@<Q1>)
{
  int v8; // w9
  int8x16_t v9; // q0
  int v10; // w10
  int8x16_t v11; // q1
  int v12; // w11
  __int64 v13; // x14
  int8x16_t *v14; // x16
  float32x4_t *v15; // x17
  int v16; // w4
  int8x16_t v17; // q2
  int8x16_t v18; // q4
  int8x16_t v19; // q5
  int8x16_t v20; // q6
  int8x16_t v21; // q7
  int32x4_t v22; // q16
  int32x4_t v23; // q5
  int v24; // w4
  float32x4_t v25; // t1
  float32x4_t v26; // q3
  float32x4_t v27; // q2
  int8x16_t *v28; // x16
  float32x4_t *v29; // x17
  int v30; // w4
  int8x16_t v31; // q3
  int8x16_t v32; // q5
  int8x16_t v33; // q7
  int8x16_t v34; // q17
  int8x16_t v35; // q19
  int32x4_t v36; // q6
  int32x4_t v37; // q16
  int v38; // w4
  float32x4_t v39; // q4
  float32x4_t v40; // q5
  float32x4_t v41; // q4
  float32x4_t v42; // q3
  float32x4_t v43; // q2
  int8x16_t *v44; // x16
  float32x4_t *v45; // x17
  int8x16_t v46; // q4
  int v47; // w4
  int8x16_t v48; // q6
  int8x16_t v49; // q16
  int8x16_t v50; // q18
  int8x16_t v51; // q20
  int32x4_t v52; // q7
  int32x4_t v53; // q17
  int v54; // w4
  float32x4_t v55; // q5
  float32x4_t v56; // q2
  float32x4_t v57; // q3
  float32x4_t v58; // q4
  int8x16_t *v59; // x16
  float32x4_t *v60; // x17
  int8x16_t v61; // q5
  int v62; // w4
  float32x4_t v63; // q21
  int8x16_t v64; // q7
  int8x16_t v65; // q17
  float32x4_t v66; // q16
  float32x4_t v67; // q18
  int8x16_t v68; // q19
  int8x16_t v69; // q21
  int32x4_t v70; // q16
  int32x4_t v71; // q18
  int v72; // w4
  float32x4_t v73; // q6
  float32x4_t v74; // q7
  float32x4_t v75; // q16
  float32x4_t v76; // q6
  float32x4_t v77; // q5
  float32x4_t v78; // q2
  float32x4_t v79; // q3
  float32x4_t v80; // q4
  int8x16_t *v81; // x16
  int8x16_t v82; // q6
  float32x4_t *v83; // x17
  int v84; // w4
  int8x16_t v85; // q16
  int8x16_t v86; // q7
  int8x16_t v87; // q19
  int8x16_t v88; // q21
  int32x4_t v89; // q17
  int32x4_t v90; // q20
  int v91; // w4
  float32x4_t v92; // q17
  float32x4_t v93; // q7
  float32x4_t v94; // q7
  float32x4_t v95; // q4
  float32x4_t v96; // q2
  float32x4_t v97; // q3
  float32x4_t v98; // q5
  float32x4_t v99; // q6
  int8x16_t *v100; // x16
  int8x16_t v101; // q7
  float32x4_t *v102; // x17
  int v103; // w4
  int8x16_t v104; // q18
  int8x16_t v105; // q17
  int8x16_t v106; // q19
  int32x4_t v107; // q21
  int8x16_t v108; // q20
  int32x4_t v109; // q22
  int v110; // w4
  float32x4_t v111; // q16
  float32x4_t v112; // q17
  float32x4_t v113; // q18
  float32x4_t v114; // q16
  float32x4_t v115; // q7
  float32x4_t v116; // q2
  float32x4_t v117; // q3
  float32x4_t v118; // q4
  float32x4_t v119; // q5
  float32x4_t v120; // q6
  int8x16_t v121; // q16
  int8x16_t *v122; // x16
  float32x4_t *v123; // x17
  int v124; // w4
  int8x16_t v125; // q18
  int8x16_t v126; // q17
  int8x16_t v127; // q19
  int8x16_t v128; // q20
  int32x4_t v129; // q21
  int32x4_t v130; // q22
  int v131; // w4
  float32x4_t v132; // q18
  float32x4_t v133; // q17
  float32x4_t v134; // q17
  float32x4_t v135; // q6
  float32x4_t v136; // q2
  float32x4_t v137; // q3
  float32x4_t v138; // q4
  float32x4_t v139; // q5
  float32x4_t v140; // q7
  float32x4_t v141; // q16
  int8x16_t v142; // q17
  int8x16_t *v143; // x16
  float32x4_t *v144; // x17
  int v145; // w4
  float32x4_t v146; // q22
  int8x16_t v147; // q20
  int8x16_t v148; // q19
  int32x4_t v149; // q23
  float32x4_t v150; // q24
  float32x4_t v151; // q27
  int8x16_t v152; // q21
  int8x16_t v153; // q22
  int32x4_t v154; // q24
  int v155; // w4
  float32x4_t v156; // q18
  float32x4_t v157; // q19
  float32x4_t v158; // q20
  float32x4_t v159; // q18

  v9.n128_u64[1] = -1;
  v9.n128_u64[0] = -1;
  v10 = 0;
  v11 = vmvnq_s8(a8);
  v12 = a6 >> 2;
  v13 = a5;
  while ( v10 < a3 )
  {
    switch ( a7 )
    {
      case 0:
      case 1:
      case 2:
      case 3:
        v14 = a2;
        v15 = a4;
        v16 = a6 >> 2;
        v17 = vandq_s8(*result, v9);
        if ( v12 )
        {
          do
          {
            v18 = vmulq_f32(*v15, v17);
            v19 = vmulq_f32(v15[8], v17);
            v20 = vmulq_f32(v15[16], v17);
            v21 = vmulq_f32(v15[24], v17);
            v22 = vextq_s8(v18, vextq_s8(v19, v18, 8u), 8u);
            v18.n128_u64[1] = v19.n128_u64[0];
            v23 = vextq_s8(v20, vextq_s8(v21, v20, 8u), 8u);
            v20.n128_u64[1] = v21.n128_u64[0];
            --v16;
            *v14 = vaddq_f32(
                     vuzp1q_s32(v18, v20),
                     vaddq_f32(
                       vaddq_f32(vaddq_f32(vandq_s8(*v14, v11), vuzp2q_s32(v18, v20)), vuzp1q_s32(v22, v23)),
                       vuzp2q_s32(v22, v23)));
            ++v14;
            v15 += 32;
          }
          while ( v16 );
        }
        v24 = a6 & 3;
        if ( (a6 & 3) != 0 )
        {
          do
          {
            v25 = *v15;
            v15 += 8;
            --v24;
            v26 = vmulq_f32(v25, v17);
            v14->n128_f32[0] = vaddv_f32((float32x2_t)vpaddq_f32(v26, v26).n128_u64[0])
                             + COERCE_FLOAT(v14->n128_u32[0] & v8);
            v14 = (int8x16_t *)((char *)v14 + 4);
          }
          while ( v24 );
        }
        goto LABEL_43;
      case 4:
      case 5:
      case 6:
      case 7:
        v27 = *result;
        v28 = a2;
        v29 = a4;
        v30 = a6 >> 2;
        v31 = vandq_s8(result[1], v9);
        if ( v12 )
        {
          do
          {
            v32 = vmlaq_f32(vmulq_f32(*v29, v27), v31, v29[1]);
            v33 = vmlaq_f32(vmulq_f32(v29[8], v27), v31, v29[9]);
            v34 = vmlaq_f32(vmulq_f32(v29[16], v27), v31, v29[17]);
            v35 = vmlaq_f32(vmulq_f32(v29[24], v27), v31, v29[25]);
            v36 = vextq_s8(v32, vextq_s8(v33, v32, 8u), 8u);
            v37 = vextq_s8(v34, vextq_s8(v35, v34, 8u), 8u);
            v32.n128_u64[1] = v33.n128_u64[0];
            v34.n128_u64[1] = v35.n128_u64[0];
            --v30;
            *v28 = vaddq_f32(
                     vuzp1q_s32(v32, v34),
                     vaddq_f32(
                       vaddq_f32(vaddq_f32(vandq_s8(*v28, v11), vuzp2q_s32(v32, v34)), vuzp1q_s32(v36, v37)),
                       vuzp2q_s32(v36, v37)));
            ++v28;
            v29 += 32;
          }
          while ( v30 );
        }
        v38 = a6 & 3;
        if ( (a6 & 3) != 0 )
        {
          do
          {
            v39 = *v29;
            v40 = v29[1];
            v29 += 8;
            --v38;
            v41 = vmlaq_f32(vmulq_f32(v39, v27), v31, v40);
            v28->n128_f32[0] = vaddv_f32((float32x2_t)vpaddq_f32(v41, v41).n128_u64[0])
                             + COERCE_FLOAT(v28->n128_u32[0] & v8);
            v28 = (int8x16_t *)((char *)v28 + 4);
          }
          while ( v38 );
        }
        goto LABEL_43;
      case 8:
      case 9:
      case 10:
      case 11:
        v42 = result[1];
        v43 = *result;
        v44 = a2;
        v45 = a4;
        v46 = vandq_s8(result[2], v9);
        v47 = a6 >> 2;
        if ( v12 )
        {
          do
          {
            v48 = vmlaq_f32(vmlaq_f32(vmulq_f32(*v45, v43), v42, v45[1]), v46, v45[2]);
            v49 = vmlaq_f32(vmlaq_f32(vmulq_f32(v45[8], v43), v42, v45[9]), v46, v45[10]);
            v50 = vmlaq_f32(vmlaq_f32(vmulq_f32(v45[16], v43), v42, v45[17]), v46, v45[18]);
            v51 = vmlaq_f32(vmlaq_f32(vmulq_f32(v45[24], v43), v42, v45[25]), v46, v45[26]);
            v52 = vextq_s8(v48, vextq_s8(v49, v48, 8u), 8u);
            v53 = vextq_s8(v50, vextq_s8(v51, v50, 8u), 8u);
            v48.n128_u64[1] = v49.n128_u64[0];
            v50.n128_u64[1] = v51.n128_u64[0];
            --v47;
            *v44 = vaddq_f32(
                     vuzp1q_s32(v48, v50),
                     vaddq_f32(
                       vaddq_f32(vaddq_f32(vandq_s8(*v44, v11), vuzp2q_s32(v48, v50)), vuzp1q_s32(v52, v53)),
                       vuzp2q_s32(v52, v53)));
            ++v44;
            v45 += 32;
          }
          while ( v47 );
        }
        v54 = a6 & 3;
        if ( (a6 & 3) != 0 )
        {
          do
          {
            --v54;
            v55 = vmlaq_f32(vmlaq_f32(vmulq_f32(*v45, v43), v42, v45[1]), v46, v45[2]);
            v44->n128_f32[0] = vaddv_f32((float32x2_t)vpaddq_f32(v55, v55).n128_u64[0])
                             + COERCE_FLOAT(v44->n128_u32[0] & v8);
            v44 = (int8x16_t *)((char *)v44 + 4);
            v45 += 8;
          }
          while ( v54 );
        }
        goto LABEL_43;
      case 12:
      case 13:
      case 14:
      case 15:
        v56 = result[2];
        v57 = *result;
        v58 = result[1];
        v59 = a2;
        v60 = a4;
        v61 = vandq_s8(result[3], v9);
        v62 = a6 >> 2;
        if ( v12 )
        {
          do
          {
            v63 = vmlaq_f32(vmulq_f32(v60[24], v57), v58, v60[25]);
            v64 = vmlaq_f32(vmlaq_f32(vmlaq_f32(vmulq_f32(*v60, v57), v58, v60[1]), v56, v60[2]), v61, v60[3]);
            v65 = vmlaq_f32(vmlaq_f32(vmlaq_f32(vmulq_f32(v60[8], v57), v58, v60[9]), v56, v60[10]), v61, v60[11]);
            --v62;
            v67 = v60[26];
            v66 = v60[27];
            v68 = vmlaq_f32(vmlaq_f32(vmlaq_f32(vmulq_f32(v60[16], v57), v58, v60[17]), v56, v60[18]), v61, v60[19]);
            v60 += 32;
            v69 = vmlaq_f32(vmlaq_f32(v63, v56, v67), v61, v66);
            v70 = vextq_s8(v64, vextq_s8(v65, v64, 8u), 8u);
            v71 = vextq_s8(v68, vextq_s8(v69, v68, 8u), 8u);
            v64.n128_u64[1] = v65.n128_u64[0];
            v68.n128_u64[1] = v69.n128_u64[0];
            *v59 = vaddq_f32(
                     vuzp1q_s32(v64, v68),
                     vaddq_f32(
                       vaddq_f32(vaddq_f32(vandq_s8(*v59, v11), vuzp2q_s32(v64, v68)), vuzp1q_s32(v70, v71)),
                       vuzp2q_s32(v70, v71)));
            ++v59;
          }
          while ( v62 );
        }
        v72 = a6 & 3;
        if ( (a6 & 3) != 0 )
        {
          do
          {
            --v72;
            v73 = vmlaq_f32(vmulq_f32(*v60, v57), v58, v60[1]);
            v75 = v60[2];
            v74 = v60[3];
            v60 += 8;
            v76 = vmlaq_f32(vmlaq_f32(v73, v56, v75), v61, v74);
            v59->n128_f32[0] = vaddv_f32((float32x2_t)vpaddq_f32(v76, v76).n128_u64[0])
                             + COERCE_FLOAT(v59->n128_u32[0] & v8);
            v59 = (int8x16_t *)((char *)v59 + 4);
          }
          while ( v72 );
        }
        goto LABEL_43;
      case 16:
      case 17:
      case 18:
      case 19:
        v77 = result[3];
        v78 = *result;
        v79 = result[1];
        v80 = result[2];
        v81 = a2;
        v82 = vandq_s8(result[4], v9);
        v83 = a4;
        v84 = a6 >> 2;
        if ( v12 )
        {
          do
          {
            --v84;
            v85 = vmlaq_f32(
                    vmlaq_f32(vmlaq_f32(vmlaq_f32(vmulq_f32(*v83, v78), v79, v83[1]), v80, v83[2]), v77, v83[3]),
                    v82,
                    v83[4]);
            v86 = vmlaq_f32(
                    vmlaq_f32(vmlaq_f32(vmlaq_f32(vmulq_f32(v83[8], v78), v79, v83[9]), v80, v83[10]), v77, v83[11]),
                    v82,
                    v83[12]);
            v87 = vmlaq_f32(
                    vmlaq_f32(vmlaq_f32(vmlaq_f32(vmulq_f32(v83[16], v78), v79, v83[17]), v80, v83[18]), v77, v83[19]),
                    v82,
                    v83[20]);
            v88 = vmlaq_f32(
                    vmlaq_f32(vmlaq_f32(vmlaq_f32(vmulq_f32(v83[24], v78), v79, v83[25]), v80, v83[26]), v77, v83[27]),
                    v82,
                    v83[28]);
            v89 = vextq_s8(v85, vextq_s8(v86, v85, 8u), 8u);
            v90 = vextq_s8(v87, vextq_s8(v88, v87, 8u), 8u);
            v85.n128_u64[1] = v86.n128_u64[0];
            v87.n128_u64[1] = v88.n128_u64[0];
            *v81 = vaddq_f32(
                     vuzp1q_s32(v85, v87),
                     vaddq_f32(
                       vaddq_f32(vaddq_f32(vandq_s8(*v81, v11), vuzp2q_s32(v85, v87)), vuzp1q_s32(v89, v90)),
                       vuzp2q_s32(v89, v90)));
            ++v81;
            v83 += 32;
          }
          while ( v84 );
        }
        v91 = a6 & 3;
        if ( (a6 & 3) != 0 )
        {
          do
          {
            --v91;
            v92 = v83[4];
            v93 = vmlaq_f32(vmlaq_f32(vmlaq_f32(vmulq_f32(*v83, v78), v79, v83[1]), v80, v83[2]), v77, v83[3]);
            v83 += 8;
            v94 = vmlaq_f32(v93, v82, v92);
            v81->n128_f32[0] = vaddv_f32((float32x2_t)vpaddq_f32(v94, v94).n128_u64[0])
                             + COERCE_FLOAT(v81->n128_u32[0] & v8);
            v81 = (int8x16_t *)((char *)v81 + 4);
          }
          while ( v91 );
        }
        goto LABEL_43;
      case 20:
      case 21:
      case 22:
      case 23:
        v95 = result[4];
        v96 = *result;
        v97 = result[1];
        v98 = result[2];
        v99 = result[3];
        v100 = a2;
        v101 = vandq_s8(result[5], v9);
        v102 = a4;
        v103 = a6 >> 2;
        if ( v12 )
        {
          do
          {
            v104 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(vmlaq_f32(vmlaq_f32(vmulq_f32(*v102, v96), v97, v102[1]), v98, v102[2]), v99, v102[3]),
                       v95,
                       v102[4]),
                     v101,
                     v102[5]);
            --v103;
            v105 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(vmlaq_f32(vmulq_f32(v102[8], v96), v97, v102[9]), v98, v102[10]),
                         v99,
                         v102[11]),
                       v95,
                       v102[12]),
                     v101,
                     v102[13]);
            v106 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(vmlaq_f32(vmulq_f32(v102[16], v96), v97, v102[17]), v98, v102[18]),
                         v99,
                         v102[19]),
                       v95,
                       v102[20]),
                     v101,
                     v102[21]);
            v107 = vextq_s8(v104, vextq_s8(v105, v104, 8u), 8u);
            v108 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(vmlaq_f32(vmulq_f32(v102[24], v96), v97, v102[25]), v98, v102[26]),
                         v99,
                         v102[27]),
                       v95,
                       v102[28]),
                     v101,
                     v102[29]);
            v109 = vextq_s8(v106, vextq_s8(v108, v106, 8u), 8u);
            v104.n128_u64[1] = v105.n128_u64[0];
            v106.n128_u64[1] = v108.n128_u64[0];
            *v100 = vaddq_f32(
                      vuzp1q_s32(v104, v106),
                      vaddq_f32(
                        vaddq_f32(vaddq_f32(vandq_s8(*v100, v11), vuzp2q_s32(v104, v106)), vuzp1q_s32(v107, v109)),
                        vuzp2q_s32(v107, v109)));
            ++v100;
            v102 += 32;
          }
          while ( v103 );
        }
        v110 = a6 & 3;
        if ( (a6 & 3) != 0 )
        {
          do
          {
            --v110;
            v111 = vmlaq_f32(vmlaq_f32(vmlaq_f32(vmulq_f32(*v102, v96), v97, v102[1]), v98, v102[2]), v99, v102[3]);
            v113 = v102[4];
            v112 = v102[5];
            v102 += 8;
            v114 = vmlaq_f32(vmlaq_f32(v111, v95, v113), v101, v112);
            v100->n128_f32[0] = vaddv_f32((float32x2_t)vpaddq_f32(v114, v114).n128_u64[0])
                              + COERCE_FLOAT(v100->n128_u32[0] & v8);
            v100 = (int8x16_t *)((char *)v100 + 4);
          }
          while ( v110 );
        }
        goto LABEL_43;
      case 24:
      case 25:
      case 26:
      case 27:
        v115 = result[5];
        v116 = *result;
        v117 = result[1];
        v118 = result[2];
        v119 = result[3];
        v120 = result[4];
        v121 = vandq_s8(result[6], v9);
        v122 = a2;
        v123 = a4;
        v124 = a6 >> 2;
        if ( v12 )
        {
          do
          {
            --v124;
            v125 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(
                           vmlaq_f32(vmlaq_f32(vmulq_f32(*v123, v116), v117, v123[1]), v118, v123[2]),
                           v119,
                           v123[3]),
                         v120,
                         v123[4]),
                       v115,
                       v123[5]),
                     v121,
                     v123[6]);
            v126 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(
                           vmlaq_f32(vmlaq_f32(vmulq_f32(v123[8], v116), v117, v123[9]), v118, v123[10]),
                           v119,
                           v123[11]),
                         v120,
                         v123[12]),
                       v115,
                       v123[13]),
                     v121,
                     v123[14]);
            v127 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(
                           vmlaq_f32(vmlaq_f32(vmulq_f32(v123[16], v116), v117, v123[17]), v118, v123[18]),
                           v119,
                           v123[19]),
                         v120,
                         v123[20]),
                       v115,
                       v123[21]),
                     v121,
                     v123[22]);
            v128 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(
                           vmlaq_f32(vmlaq_f32(vmulq_f32(v123[24], v116), v117, v123[25]), v118, v123[26]),
                           v119,
                           v123[27]),
                         v120,
                         v123[28]),
                       v115,
                       v123[29]),
                     v121,
                     v123[30]);
            v129 = vextq_s8(v125, vextq_s8(v126, v125, 8u), 8u);
            v130 = vextq_s8(v127, vextq_s8(v128, v127, 8u), 8u);
            v125.n128_u64[1] = v126.n128_u64[0];
            v127.n128_u64[1] = v128.n128_u64[0];
            *v122 = vaddq_f32(
                      vuzp1q_s32(v125, v127),
                      vaddq_f32(
                        vaddq_f32(vaddq_f32(vandq_s8(*v122, v11), vuzp2q_s32(v125, v127)), vuzp1q_s32(v129, v130)),
                        vuzp2q_s32(v129, v130)));
            ++v122;
            v123 += 32;
          }
          while ( v124 );
        }
        v131 = a6 & 3;
        if ( (a6 & 3) != 0 )
        {
          do
          {
            --v131;
            v132 = v123[6];
            v133 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(vmlaq_f32(vmulq_f32(*v123, v116), v117, v123[1]), v118, v123[2]),
                         v119,
                         v123[3]),
                       v120,
                       v123[4]),
                     v115,
                     v123[5]);
            v123 += 8;
            v134 = vmlaq_f32(v133, v121, v132);
            v122->n128_f32[0] = vaddv_f32((float32x2_t)vpaddq_f32(v134, v134).n128_u64[0])
                              + COERCE_FLOAT(v122->n128_u32[0] & v8);
            v122 = (int8x16_t *)((char *)v122 + 4);
          }
          while ( v131 );
        }
        goto LABEL_43;
      case 28:
      case 29:
      case 30:
      case 31:
        v135 = result[6];
        v136 = *result;
        v137 = result[1];
        v138 = result[2];
        v139 = result[3];
        v140 = result[4];
        v141 = result[5];
        v142 = vandq_s8(result[7], v9);
        v143 = a2;
        v144 = a4;
        v145 = a6 >> 2;
        if ( v12 )
        {
          do
          {
            --v145;
            v146 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(vmlaq_f32(vmulq_f32(v144[24], v136), v137, v144[25]), v138, v144[26]),
                         v139,
                         v144[27]),
                       v140,
                       v144[28]),
                     v141,
                     v144[29]);
            v147 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(
                           vmlaq_f32(
                             vmlaq_f32(vmlaq_f32(vmulq_f32(*v144, v136), v137, v144[1]), v138, v144[2]),
                             v139,
                             v144[3]),
                           v140,
                           v144[4]),
                         v141,
                         v144[5]),
                       v135,
                       v144[6]),
                     v142,
                     v144[7]);
            v148 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(
                           vmlaq_f32(
                             vmlaq_f32(vmlaq_f32(vmulq_f32(v144[8], v136), v137, v144[9]), v138, v144[10]),
                             v139,
                             v144[11]),
                           v140,
                           v144[12]),
                         v141,
                         v144[13]),
                       v135,
                       v144[14]),
                     v142,
                     v144[15]);
            v149 = vextq_s8(v147, vextq_s8(v148, v147, 8u), 8u);
            v150 = v144[30];
            v151 = v144[31];
            v152 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(
                           vmlaq_f32(
                             vmlaq_f32(vmlaq_f32(vmulq_f32(v144[16], v136), v137, v144[17]), v138, v144[18]),
                             v139,
                             v144[19]),
                           v140,
                           v144[20]),
                         v141,
                         v144[21]),
                       v135,
                       v144[22]),
                     v142,
                     v144[23]);
            v147.n128_u64[1] = v148.n128_u64[0];
            v144 += 32;
            v153 = vmlaq_f32(vmlaq_f32(v146, v135, v150), v142, v151);
            v154 = vextq_s8(v152, vextq_s8(v153, v152, 8u), 8u);
            v152.n128_u64[1] = v153.n128_u64[0];
            *v143 = vaddq_f32(
                      vuzp1q_s32(v147, v152),
                      vaddq_f32(
                        vaddq_f32(vaddq_f32(vandq_s8(*v143, v11), vuzp2q_s32(v147, v152)), vuzp1q_s32(v149, v154)),
                        vuzp2q_s32(v149, v154)));
            ++v143;
          }
          while ( v145 );
        }
        v155 = a6 & 3;
        if ( (a6 & 3) != 0 )
        {
          do
          {
            --v155;
            v156 = vmlaq_f32(
                     vmlaq_f32(
                       vmlaq_f32(
                         vmlaq_f32(vmlaq_f32(vmulq_f32(*v144, v136), v137, v144[1]), v138, v144[2]),
                         v139,
                         v144[3]),
                       v140,
                       v144[4]),
                     v141,
                     v144[5]);
            v157 = v144[6];
            v158 = v144[7];
            v144 += 8;
            v159 = vmlaq_f32(vmlaq_f32(v156, v135, v157), v142, v158);
            v143->n128_f32[0] = vaddv_f32((float32x2_t)vpaddq_f32(v159, v159).n128_u64[0])
                              + COERCE_FLOAT(v143->n128_u32[0] & v8);
            v143 = (int8x16_t *)((char *)v143 + 4);
          }
          while ( v155 );
        }
LABEL_43:
        a2 = (int8x16_t *)((char *)a2 + 4 * a6);
        break;
      default:
        break;
    }
    result = (int8x16_t *)((char *)result + 4 * v13);
    ++v10;
  }
  return result;
}