stmdb   sp!, {r4, r5, r6, r7, r8, r9, r10, r11}
    ldmia   r0, {r1, r2, r3, r4, r5, r6, r7, r8}
    sub r9, r1, r3  ;t3 = z[0].re - z[1].re;
    add r10, r3, r1 ;t1 = z[0].re + z[1].re;
                ;r1, r3 now free



    sub r1, r6, r8  ;t7 = z[2].im - z[3].im;
    add r3, r6, r8  ;t5 = z[2].im + z[3].im;
                ;r6, r8 now free

    sub r11, r7, r5 ;t8 = z[3].re - z[2].re;
    add r12, r7, r5 ;t6 = z[3].re + z[2].re;
                ;r7, r5 now free


    sub r5, r2, r4  ;t4 = z[0].im - z[1].im
    add r7, r2, r4  ;t2 = z[0].im + z[1].im;
                ;r2, r4 now free

    ;have r2,r4,r6,r8 free

    sub r8, r5, r11     ;r8 = t4 - t8;
    sub r6, r7, r3      ;r6 = t2 - t5;
    add r4, r5, r11     ;r4 = t4 + t8;
    add r2, r7, r3      ;r2 = t2 + t5;
    ;have r3, r7, r5, r11 free

    sub r5, r10, r12    ;r5 = t1 - t6;
    sub r7, r9, r1      ;r7 = t3 - t7;
    add r3, r9, r1      ;r3 = t3 + t7;
    add r1, r10, r12    ;r1 = t1 + t6;

    stmia   r0, {r1, r2, r3, r4, r5, r6, r7, r8}
    ldmia   sp!, {r4, r5, r6, r7, r8, r9, r10, r11}