tustvold commented on code in PR #2781:
URL: https://github.com/apache/arrow-rs/pull/2781#discussion_r980409068
##########
arrow-buffer/src/bigint.rs:
##########
@@ -0,0 +1,253 @@
+use num::BigInt;
+use std::cmp::Ordering;
+
+/// A signed 256-bit integer
+#[allow(non_camel_case_types)]
+#[derive(Copy, Clone, Default, Eq, PartialEq, Hash)]
+pub struct i256([u8; 32]);
+
+impl std::fmt::Debug for i256 {
+ fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+ write!(f, "{}", self)
+ }
+}
+
+impl std::fmt::Display for i256 {
+ fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+ write!(f, "{}", BigInt::from_signed_bytes_le(&self.0))
+ }
+}
+
+impl PartialOrd for i256 {
+ fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
+ Some(self.cmp(other))
+ }
+}
+
+impl Ord for i256 {
+ fn cmp(&self, other: &Self) -> Ordering {
+ // This is 25x faster than using a variable length encoding such
+ // as BigInt as it avoids allocation and branching
+ let (left_low, left_high) = self.as_parts();
+ let (right_low, right_high) = other.as_parts();
+ left_high.cmp(&right_high).then(left_low.cmp(&right_low))
+ }
+}
+
+impl i256 {
+ /// Create an integer value from its representation as a byte array in
little-endian.
+ #[inline]
+ pub fn from_le_bytes(b: [u8; 32]) -> Self {
+ Self(b)
+ }
+
+ /// Return the memory representation of this integer as a byte array in
little-endian byte order.
+ #[inline]
+ pub fn to_le_bytes(self) -> [u8; 32] {
+ self.0
+ }
+
+ /// Returns this i256 as a low u128 and high i128
+ #[inline]
+ fn as_parts(self) -> (u128, i128) {
+ (
+ u128::from_le_bytes(self.0[0..16].try_into().unwrap()),
+ i128::from_le_bytes(self.0[16..32].try_into().unwrap()),
+ )
+ }
+
+ /// Create an i256 from the provided low and high i128
+ #[inline]
+ fn from_parts(low: u128, high: i128) -> Self {
+ let mut t = [0; 32];
+ let t_low: &mut [u8; 16] = (&mut t[0..16]).try_into().unwrap();
+ *t_low = low.to_le_bytes();
+ let t_high: &mut [u8; 16] = (&mut t[16..32]).try_into().unwrap();
+ *t_high = high.to_le_bytes();
+ Self(t)
+ }
+
+ /// Create an i256 from the provided [`BigInt`] returning a bool indicating
+ /// if overflow occurred
+ fn from_bigint_with_overflow(v: BigInt) -> (Self, bool) {
+ let v_bytes = v.to_signed_bytes_le();
+ match v_bytes.len().cmp(&32) {
+ Ordering::Less => {
+ let mut bytes = if num::Signed::is_negative(&v) {
+ [255_u8; 32]
+ } else {
+ [0; 32]
+ };
+
bytes[0..v_bytes.len()].copy_from_slice(&v_bytes[..v_bytes.len()]);
+ (Self(bytes), false)
+ }
+ Ordering::Equal => (Self(v_bytes.try_into().unwrap()), false),
+ Ordering::Greater => (Self(v_bytes[..32].try_into().unwrap()),
true),
+ }
+ }
+
+ /// Performs wrapping addition
+ #[inline]
+ pub fn wrapping_add(self, other: Self) -> Self {
+ let (left_low, left_high) = self.as_parts();
+ let (right_low, right_high) = other.as_parts();
+
+ let (low, carry) = left_low.overflowing_add(right_low);
+ let high = left_high.wrapping_add(right_high).wrapping_add(carry as _);
+ Self::from_parts(low, high)
+ }
+
+ /// Performs checked addition
+ #[inline]
+ pub fn checked_add(self, other: Self) -> Option<Self> {
+ let (left_low, left_high) = self.as_parts();
+ let (right_low, right_high) = other.as_parts();
+
+ let (low, carry) = left_low.overflowing_add(right_low);
+ let high = left_high.checked_add(right_high)?.checked_add(carry as _)?;
+ Some(Self::from_parts(low, high))
+ }
+
+ /// Performs wrapping addition
+ #[inline]
+ pub fn wrapping_sub(self, other: Self) -> Self {
+ let (left_low, left_high) = self.as_parts();
+ let (right_low, right_high) = other.as_parts();
+
+ let (low, carry) = left_low.overflowing_sub(right_low);
+ let high = left_high.wrapping_sub(right_high).wrapping_sub(carry as _);
+ Self::from_parts(low, high)
+ }
+
+ /// Performs checked addition
+ #[inline]
+ pub fn checked_sub(self, other: Self) -> Option<Self> {
+ let (left_low, left_high) = self.as_parts();
+ let (right_low, right_high) = other.as_parts();
+
+ let (low, carry) = left_low.overflowing_sub(right_low);
+ let high = left_high.checked_sub(right_high)?.checked_sub(carry as _)?;
+ Some(Self::from_parts(low, high))
+ }
+
+ /// Performs wrapping multiplication
+ #[inline]
+ pub fn wrapping_mul(self, other: Self) -> Self {
+ let l = BigInt::from_signed_bytes_le(&self.0);
+ let r = BigInt::from_signed_bytes_le(&other.0);
+ Self::from_bigint_with_overflow(l * r).0
+ }
+
+ /// Performs checked multiplication
+ #[inline]
+ pub fn checked_mul(self, other: Self) -> Option<Self> {
+ let l = BigInt::from_signed_bytes_le(&self.0);
+ let r = BigInt::from_signed_bytes_le(&other.0);
+ let (val, overflow) = Self::from_bigint_with_overflow(l * r);
+ (!overflow).then(|| val)
+ }
+
+ /// Performs wrapping division
+ #[inline]
+ pub fn wrapping_div(self, other: Self) -> Self {
+ let l = BigInt::from_signed_bytes_le(&self.0);
+ let r = BigInt::from_signed_bytes_le(&other.0);
+ Self::from_bigint_with_overflow(l / r).0
+ }
+
+ /// Performs checked division
+ #[inline]
+ pub fn checked_div(self, other: Self) -> Option<Self> {
+ let l = BigInt::from_signed_bytes_le(&self.0);
+ let r = BigInt::from_signed_bytes_le(&other.0);
+ let (val, overflow) = Self::from_bigint_with_overflow(l / r);
+ (!overflow).then(|| val)
+ }
+
+ /// Performs wrapping remainder
+ #[inline]
+ pub fn wrapping_rem(self, other: Self) -> Self {
+ let l = BigInt::from_signed_bytes_le(&self.0);
+ let r = BigInt::from_signed_bytes_le(&other.0);
+ Self::from_bigint_with_overflow(l % r).0
+ }
+
+ /// Performs checked division
+ #[inline]
+ pub fn checked_rem(self, other: Self) -> Option<Self> {
+ if other.0 == [0; 32] {
+ return None;
+ }
+
+ let l = BigInt::from_signed_bytes_le(&self.0);
+ let r = BigInt::from_signed_bytes_le(&other.0);
+ let (val, overflow) = Self::from_bigint_with_overflow(l % r);
+ (!overflow).then(|| val)
+ }
+}
Review Comment:
Theoretically these would all benefit from an implementation not using
BigInt, as this elides allocations and branches, but unlike the addition and
subtraction kernels these are significantly more complicated to implement. I
leave this as a potential future optimisation for someone who really cares
about decimal256 performance
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
