Big integer bit rotation - c++

Question

I want to implement unsigned left rotation in my integer class. However, because it is a template class, it can be in any size from 128-bit and goes on; so I cannot use algorithms that require a temporary of the same size because, if the type becomes big, a stack overflow will occur (specially if such function was in call chain).
So to fix such problem I minimized it to a question: what steps do I have to do to rotate a 32-bit number using only 4 bits. Well, if you think about, it a 32-bit number contains 8 groups of 4 bits each, so if the number of bits to rotate is 4 then a swap will occur between groups 0 and 4, 1 and 5, 2 and 6, 3 and 7, after which the rotation is done.
If bits to rotate less than 4 and greater than 0 then it is simple just preserve the last N bits and start shift-Or loop, e.g. suppose we have the number 0x9CE2 to left rotate it 3 bits we will do that following:
The number in little endian binary is 1001 1100 1110 0010, each nibble indexed from 0 to 3 from right to left and we will call the this number N and number of bits in one group B
[1] x <- N[3] >> 3
x <- 1001 >> 3
x <- 0100
y <- N[0] >> (B - 3)
y <- N[0] >> (4 - 3)
y <- 0010 >> 1
y <- 0001
N[0] <- (N[0] << 3) | x
N[0] <- (0010 << 3) | 0100
N[0] <- 0000 | 0100
N[0] <- 0100
[2] x <- y
x <- 0001
y <- N[1] >> (B - 3)
y <- N[1] >> (4 - 3)
y <- 1110 >> 1
y <- 0111
N[1] <- (N[1] << 3) | x
N[1] <- (1110 << 3) | 0001
N[1] <- 0000 | 0001
N[1] <- 0001
[3] x <- y
x <- 0111
y <- N[2] >> (B - 3)
y <- N[2] >> (4 - 3)
y <- 1100 >> 1
y <- 0110
N[2] <- (N[2] << 3) | x
N[2] <- (1100 << 3) | 0111
N[2] <- 0000 | 0111
N[2] <- 0111
[4] x <- y
x <- 0110
y <- N[3] >> (B - 3)
y <- N[3] >> (4 - 3)
y <- 1001 >> 1
y <- 0100
N[3] <- (N[3] << 3) | x
N[3] <- (1001 << 3) | 0110
N[3] <- 1000 | 0110
N[3] <- 1110
The result is 1110 0111 0001 0100, 0xE714 in hexadecimal, which is the right answer; and, if you try to apply it on any number with any precision, all you will need is one variable which type is the type of any element of the array forming that bignum type.
Now the real problem is when the number bits to rotate is bigger than one group or bigger than half size of the type (i.e. bigger than 4bits or 8bits in this example).
Usually, we shift bits from last element to first element and so on; but now, after shifting
the last element to first element, the result has to be relocated to new place because the number of bits to rotate is bigger than on element (i.e. > 4 bits). The start index where the shift will start is the last index (3 in this example), and for destination index we use the equation: dest_index = int(bits_count/half_bits) + 1, where half_bits is number of bits in half the number and in this example half_bits = 8, so if bits_count = 7 then dest_index = int(7/8) + 1 = 1 + 1 = 2, and that means the result of the first shift must relocated to destination index 2 -- and that is my problem, for I cannot think of a way to write an algorithm for this situation.
Thanks.

Answer 1

This will just be some hints for one way to accomplish this. You can think about making two passes.
first pass, rotate on 4 bit boundaries only
second pass, rotate on 1 bit boundaries
So, the top level pseudo code might look like:
rotate (unsigned bits) {
bits %= 32; /* only have 32 bits */
if (bits == 0) return;
rotate_nibs(bits/4);
rotate_bits(bits%4);
}
So, to rotate by 13 bits, you first rotate by 3 nibbles, then rotate by 1 bit to get your total of 13 bits of rotation.
You could avoid nibble rotation altogether if you treat your array of nibbles as a circular buffer. Then, a nibble rotation is just a matter of changing the start position in the array for the 0 index.
If you must do rotation, it can be tricky. If you are rotating an 8 item array and only want to use 1 item of storage overhead to do the rotation, then to rotate by 3 items, you might approach it like this:
orig: A B C D E F G H
step 1: A B C A E F G H rem: D
2: A B C A E F D H rem: G
3: A G C A E F D H rem: B
4: A G C A B F D H rem: E
5: A G C A B F D E rem: H
6: A G H A B F D E rem: C
7: A G H A B C D E rem: F
8: F G H A B C D E done
But, if you tried the same technique with 2, 4, or 6 item rotations, the cycle does not run through the whole array. So, you have to be aware if the rotation count and the array size has a common divisor, and make the algorithm account for that. If you step through with the 6 step rotation, some more clues fall out.
orig: A B C D E F G H
A
G
E
C cycled back to A's position
B
H
F
D done
Notice that the GCD(6,8) is 2, which means we should expect 4 iterations for each pass. Then, the rotation algorithm for an N item array could look like:
rotate (n) {
G = GCD(n, N)
for (i = 0; i < G; ++i) {
p = arr[i];
for (j = 1; j < N/G; ++j) {
swap(p, arr[(i + j*n) % N]);
}
arr[i] = p;
}
}
There is an optimization you can do to avoid the swap per iteration, that I'll leave as an exercise.

Answer 2

I suggest calling an assembly language function for the bit rotation.
Many assembly languages have better facilities for rotating bits through carry and rotating carry through bits.
Many times the assembly language is less complex than the C or C++ function.
The drawback is that you will need one instance of each assembly function for each {different} platform.