ld1         {{ v9.4s, v10.4s, v11.4s }}, [x1], #48

    fmla        v16.4s, v0.4s, v8.s[0]
    ldr         x8, [x2], #8
    fmla        v17.4s, v1.4s, v8.s[0]
    ldr         d12, [x1], #8
    fmla        v18.4s, v2.4s, v8.s[0]
    ldr         x12, [x1], #8
    fmla        v19.4s, v3.4s, v8.s[0]
    ldr         d13, [x1], #8
    fmla        v20.4s, v4.4s, v8.s[0]
    ldr         x13, [x1], #8
    fmla        v21.4s, v5.4s, v8.s[0]
    ldr         d14, [x1], #8
    fmla        v22.4s, v6.4s, v8.s[0]
    ldr         x14, [x1], #8
    fmla        v23.4s, v7.4s, v8.s[0]
    ldr         d15, [x1], #8
    fmla        v24.4s, v9.4s, v8.s[0]
    ldr         x15, [x1], #8

    ld1         {{ v0.4s, v1.4s, v2.4s, v3.4s }}, [x1], #64
    ins         v8.d[1], x8
    ld1         {{ v4.4s, v5.4s, v6.4s, v7.4s }}, [x1], #64

    fmla        v25.4s, v10.4s, v8.s[0]
    ins         v12.d[1], x12
    fmla        v26.4s, v11.4s, v8.s[0]
    ins         v13.d[1], x13
    fmla        v27.4s, v12.4s, v8.s[0]
    ins         v14.d[1], x14
    fmla        v28.4s, v13.4s, v8.s[0]
    ins         v15.d[1], x15

    ld1         {{ v9.4s, v10.4s, v11.4s, v12.4s }}, [x1], #64

    fmla        v29.4s, v14.4s, v8.s[0]
    ldr         d13, [x1], #8
    fmla        v30.4s, v15.4s, v8.s[0]
    ldr         x13, [x1], #8
    fmla        v31.4s, v0.4s, v8.s[0]
    ldr         d14, [x1], #8

    fmla        v16.4s, v1.4s, v8.s[2]
    ldr         x14, [x1], #8
    fmla        v17.4s, v2.4s, v8.s[2]
    ldr         d15, [x1], #8
    fmla        v18.4s, v3.4s, v8.s[2]
    ldr         x15, [x1], #8
    fmla        v19.4s, v4.4s, v8.s[2]

    ld1         {{ v0.4s }}, [x1], #16

    fmla        v20.4s, v5.4s, v8.s[2]
    ldr         d1, [x1], #8
    fmla        v21.4s, v6.4s, v8.s[2]
    ldr         x10, [x1], #8

    fmla        v22.4s, v7.4s, v8.s[2]

    fmla        v23.4s, v9.4s, v8.s[2]
    ins         v13.d[1], x13
    fmla        v24.4s, v10.4s, v8.s[2]
    ins         v14.d[1], x14
    fmla        v25.4s, v11.4s, v8.s[2]
    ins         v15.d[1], x15

    fmla        v26.4s, v12.4s, v8.s[2]
    prfm        pldl1keep, [x1, #1024]
    fmla        v27.4s, v13.4s, v8.s[2]
    ins         v1.d[1], x10
    fmla        v28.4s, v14.4s, v8.s[2]
    prfm        pldl1keep, [x1, #1088]
    fmla        v29.4s, v15.4s, v8.s[2]
    prfm        pldl1keep, [x1, #1152]
    fmla        v30.4s, v0.4s, v8.s[2]
    prfm        pldl1keep, [x1, #1216]
    fmla        v31.4s, v1.4s, v8.s[2]
    prfm        pldl1keep, [x2, #256]

    ld1         {{ v0.4s, v1.4s, v2.4s, v3.4s }}, [x1], #64
    ins         v8.s[0], v8.s[3]
    ld1         {{ v4.4s, v5.4s, v6.4s, v7.4s }}, [x1], #64