/**
* Not necessarily optimal... but it produces correct results (at least for int)
* If we're lucky, LLVM will recognize the pattern and produce rotate
* instructions:
* http://llvm.1065342.n5.nabble.com/rotate-td47679.html
*
* Eventually, someone should feel free to implement an llvm-specific version
*/
_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE rotate(__CLC_GENTYPE x, __CLC_GENTYPE n){
//Try to avoid extra work if someone's spinning the value through multiple
//full rotations
n = n % (__CLC_GENTYPE)__CLC_GENSIZE;
#ifdef __CLC_SCALAR
if (n > 0){
return (x << n) | (((__CLC_U_GENTYPE)x) >> (__CLC_GENSIZE - n));
} else if (n == 0){
return x;
} else {
return ( (((__CLC_U_GENTYPE)x) >> -n) | (x << (__CLC_GENSIZE + n)) );
}
#else
//XXX: There's a lot of __builtin_astype calls to cast everything to
// unsigned ... This should be improved so that if __CLC_GENTYPE==__CLC_U_GENTYPE, no
// casts are required.
__CLC_U_GENTYPE x_1 = __builtin_astype(x, __CLC_U_GENTYPE);
//XXX: Is (__CLC_U_GENTYPE >> S__CLC_GENTYPE) | (__CLC_U_GENTYPE << S__CLC_GENTYPE) legal?
// If so, then combine the amt and shifts into a single set of statements
__CLC_U_GENTYPE amt;
amt = (n < (__CLC_GENTYPE)0 ? __builtin_astype((__CLC_GENTYPE)0-n, __CLC_U_GENTYPE) : (__CLC_U_GENTYPE)0);
x_1 = (x_1 >> amt) | (x_1 << ((__CLC_U_GENTYPE)__CLC_GENSIZE - amt));
amt = (n < (__CLC_GENTYPE)0 ? (__CLC_U_GENTYPE)0 : __builtin_astype(n, __CLC_U_GENTYPE));
x_1 = (x_1 << amt) | (x_1 >> ((__CLC_U_GENTYPE)__CLC_GENSIZE - amt));
return __builtin_astype(x_1, __CLC_GENTYPE);
#endif
}