| 1 | /// Creates an unsigned division function optimized for dividing integers with the same |
| 2 | /// bitwidth as the largest operand in an asymmetrically sized division. For example, x86-64 has an |
| 3 | /// assembly instruction that can divide a 128 bit integer by a 64 bit integer if the quotient fits |
| 4 | /// in 64 bits. The 128 bit version of this algorithm would use that fast hardware division to |
| 5 | /// construct a full 128 bit by 128 bit division. |
| 6 | #[allow (unused_macros)] |
| 7 | macro_rules! impl_asymmetric { |
| 8 | ( |
| 9 | $fn:ident, // name of the unsigned division function |
| 10 | $zero_div_fn:ident, // function called when division by zero is attempted |
| 11 | $half_division:ident, // function for division of a $uX by a $uX |
| 12 | $asymmetric_division:ident, // function for division of a $uD by a $uX |
| 13 | $n_h:expr, // the number of bits in a $iH or $uH |
| 14 | $uH:ident, // unsigned integer with half the bit width of $uX |
| 15 | $uX:ident, // unsigned integer with half the bit width of $uD |
| 16 | $uD:ident // unsigned integer type for the inputs and outputs of `$fn` |
| 17 | ) => { |
| 18 | /// Computes the quotient and remainder of `duo` divided by `div` and returns them as a |
| 19 | /// tuple. |
| 20 | pub fn $fn(duo: $uD, div: $uD) -> ($uD, $uD) { |
| 21 | let n: u32 = $n_h * 2; |
| 22 | |
| 23 | let duo_lo = duo as $uX; |
| 24 | let duo_hi = (duo >> n) as $uX; |
| 25 | let div_lo = div as $uX; |
| 26 | let div_hi = (div >> n) as $uX; |
| 27 | if div_hi == 0 { |
| 28 | if div_lo == 0 { |
| 29 | $zero_div_fn() |
| 30 | } |
| 31 | if duo_hi < div_lo { |
| 32 | // `$uD` by `$uX` division with a quotient that will fit into a `$uX` |
| 33 | let (quo, rem) = unsafe { $asymmetric_division(duo, div_lo) }; |
| 34 | return (quo as $uD, rem as $uD); |
| 35 | } else { |
| 36 | // Short division using the $uD by $uX division |
| 37 | let (quo_hi, rem_hi) = $half_division(duo_hi, div_lo); |
| 38 | let tmp = unsafe { |
| 39 | $asymmetric_division((duo_lo as $uD) | ((rem_hi as $uD) << n), div_lo) |
| 40 | }; |
| 41 | return ((tmp.0 as $uD) | ((quo_hi as $uD) << n), tmp.1 as $uD); |
| 42 | } |
| 43 | } |
| 44 | |
| 45 | // This has been adapted from |
| 46 | // https://www.codeproject.com/tips/785014/uint-division-modulus which was in turn |
| 47 | // adapted from Hacker's Delight. This is similar to the two possibility algorithm |
| 48 | // in that it uses only more significant parts of `duo` and `div` to divide a large |
| 49 | // integer with a smaller division instruction. |
| 50 | let div_lz = div_hi.leading_zeros(); |
| 51 | let div_extra = n - div_lz; |
| 52 | let div_sig_n = (div >> div_extra) as $uX; |
| 53 | let tmp = unsafe { $asymmetric_division(duo >> 1, div_sig_n) }; |
| 54 | |
| 55 | let mut quo = tmp.0 >> ((n - 1) - div_lz); |
| 56 | if quo != 0 { |
| 57 | quo -= 1; |
| 58 | } |
| 59 | |
| 60 | // Note that this is a full `$uD` multiplication being used here |
| 61 | let mut rem = duo - (quo as $uD).wrapping_mul(div); |
| 62 | if div <= rem { |
| 63 | quo += 1; |
| 64 | rem -= div; |
| 65 | } |
| 66 | return (quo as $uD, rem); |
| 67 | } |
| 68 | }; |
| 69 | } |
| 70 | |