asm volatile(
                "xorl %%esi, %%esi                      \n\t"
                "leal "MANGLE(bit_reverse_512)", %%eax  \n\t"
                "movl $1008, %%edi                      \n\t"
                "pushl %%ebp                            \n\t" //use ebp without telling gcc
                ".balign 16                             \n\t"
                "1:                                     \n\t"
                "movlps (%0, %%esi), %%xmm0             \n\t" // XXXI
                "movhps 8(%0, %%edi), %%xmm0            \n\t" // RXXI
                "movlps 8(%0, %%esi), %%xmm1            \n\t" // XXXi
                "movhps (%0, %%edi), %%xmm1             \n\t" // rXXi
                "shufps $0x33, %%xmm1, %%xmm0           \n\t" // irIR
                "movaps "MANGLE(sseSinCos1c)"(%%esi), %%xmm2\n\t"
                "mulps %%xmm0, %%xmm2                   \n\t"
                "shufps $0xB1, %%xmm0, %%xmm0           \n\t" // riRI
                "mulps "MANGLE(sseSinCos1d)"(%%esi), %%xmm0\n\t"
                "subps %%xmm0, %%xmm2                   \n\t"
                "movzbl (%%eax), %%edx                  \n\t"
                "movzbl 1(%%eax), %%ebp                 \n\t"
                "movlps %%xmm2, (%1, %%edx,8)           \n\t"
                "movhps %%xmm2, (%1, %%ebp,8)           \n\t"
                "addl $16, %%esi                        \n\t"
                "addl $2, %%eax                         \n\t" // avoid complex addressing for P4 crap
                "subl $16, %%edi                        \n\t"
                " jnc 1b                                \n\t"
                "popl %%ebp                             \n\t"//no we didnt touch ebp *g*
                :: "b" (data), "c" (buf)
                : "%esi", "%edi", "%eax", "%edx"
        );