// Start of first stage with input fetching (input is assumed to be in bit
//reversed order)
// Input is assumed to be scaled by four. The outputis also scaled by four by
//ASR option
    R0 = [I0++];            // x0
    R3 = [I0++];            // x2
    R1 = [I0++];            // x1
    LSETUP(STAGE1_ST,STAGE1_END) LC0 = P1;
                            // There are N/4 butterflies in first stage
STAGE1_ST:
        R4 = R0 +|+ R3 , R5 = R0 -|- R3(ASR) || R3 = [I0++] || [I1++] = R5;
                            // A = x0 +|+ x2, B = x0 -|- x2, fetch x3, store y1
                            // of prev. bf
        R6 = R1 +|+ R3 , R1 = R1 -|- R3 (ASR,CO) || R0 = [I0++] || [I1++] = R6;
                            // C = x1 +|+ x3, D = x1 -|- x3(CO), fetch x0, store
                            // y2 of prev. bf
        R4 = R4 +|+ R6 , R6 = R4 -|- R6(ASR) || R3 = [I0++] || [I1++] = R7;
                            // y0 = A +|+ C, y2 = A -|- C, fetch x2, store y3 of
                            // prev. bf
STAGE1_END:
        R7 = R5 +|- R1 , R5 = R5 -|+ R1(ASR) || R1 = [I0++] || [I1++] = R4;
                            // y3 = B +|- D, y1 = B -|+ D, fetch x3, Store y0 of
                            // this bf
    [I1++] = R5;            // Store y1 of last butterfly
    [I1++] = R6;            // Store y2 of last butterfly
    [I1++] = R7;            // Store y3 of last butterfly
                            // End of first stage with input fetching by bit
                            // reversal