user0/rust
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Speed up SipHasher128.

Browse source 
The current code in `SipHasher128::short_write` is inefficient. It uses
`u8to64_le` (which is complex and slow) to extract just the right number of
bytes of the input into a u64 and pad the result with zeroes. It then
left-shifts that value in order to bitwise-OR it with `self.tail`.

For example, imagine we have a u32 input 0xIIHH_GGFF and only need three bytes
to fill up `self.tail`. The current code uses `u8to64_le` to construct
0x0000_0000_00HH_GGFF, which is just 0xIIHH_GGFF with the 0xII removed and
zero-extended to a u64. The code then left-shifts that value by five bytes --
discarding the 0x00 byte that replaced the 0xII byte! -- to give
0xHHGG_FF00_0000_0000. It then then ORs that value with self.tail.

There's a much simpler way to do it: zero-extend to u64 first, then left shift.
E.g. 0xIIHH_GGFF is zero-extended to 0x0000_0000_IIHH_GGFF, and then
left-shifted to 0xHHGG_FF00_0000_0000. We don't have to take time to exclude
the unneeded 0xII byte, because it just gets shifted out anyway! It also avoids
multiple occurrences of `unsafe`.

There's a similar story with the setting of `self.tail` at the method's end.
The current code uses `u8to64_le` to extract the remaining part of the input,
but the same effect can be achieved more quickly with a right shift on the
zero-extended input.

All that works on little-endian. It doesn't work for big-endian, but we
can just do a `to_le` before calling `short_write` and then it works.

This commit changes `SipHasher128` to use the simpler shift-based approach. The
code is also smaller, which means that `short_write` is now inlined where
previously it wasn't, which makes things faster again. This gives big
speed-ups for all incremental builds, especially "baseline" incremental
builds.
This commit is contained in:
Nicholas Nethercote 2020-02-06 13:04:51 +11:00
parent a19edd6b16
commit f8a02864af
1 changed files with 72 additions and 39 deletions
Download patch file Download diff file Expand all files Collapse all files

111
src/librustc_data_structures/sip128.rs
View file

@ -4,7 +4,6 @@ use std::cmp;
use std::hash::Hasher;
use std::mem;
use std::ptr;
use std::slice;
#[cfg(test)]
mod tests;
@ -122,42 +121,76 @@ impl SipHasher128 {
self.state.v1 ^= 0xee;
}
// Specialized write function that is only valid for buffers with len <= 8.
// It's used to force inlining of write_u8 and write_usize, those would normally be inlined
// except for composite types (that includes slices and str hashing because of delimiter).
// Without this extra push the compiler is very reluctant to inline delimiter writes,
// degrading performance substantially for the most common use cases.
// A specialized write function for values with size <= 8.
//
// The hashing of multi-byte integers depends on endianness. E.g.:
// - little-endian: `write_u32(0xDDCCBBAA)` == `write([0xAA, 0xBB, 0xCC, 0xDD])`
// - big-endian: `write_u32(0xDDCCBBAA)` == `write([0xDD, 0xCC, 0xBB, 0xAA])`
//
// This function does the right thing for little-endian hardware. On
// big-endian hardware `x` must be byte-swapped first to give the right
// behaviour. After any byte-swapping, the input must be zero-extended to
// 64-bits. The caller is responsible for the byte-swapping and
// zero-extension.
#[inline]
fn short_write(&mut self, msg: &[u8]) {
debug_assert!(msg.len() <= 8);
let length = msg.len();
self.length += length;
fn short_write<T>(&mut self, _x: T, x: u64) {
let size = mem::size_of::<T>();
self.length += size;
// The original number must be zero-extended, not sign-extended.
debug_assert!(if size < 8 { x >> (8 * size) == 0 } else { true });
// The number of bytes needed to fill `self.tail`.
let needed = 8 - self.ntail;
let fill = cmp::min(length, needed);
if fill == 8 {
self.tail = unsafe { load_int_le!(msg, 0, u64) };
} else {
self.tail |= unsafe { u8to64_le(msg, 0, fill) } << (8 * self.ntail);
if length < needed {
self.ntail += length;
return;
}
// SipHash parses the input stream as 8-byte little-endian integers.
// Inputs are put into `self.tail` until 8 bytes of data have been
// collected, and then that word is processed.
//
// For example, imagine that `self.tail` is 0x0000_00EE_DDCC_BBAA,
// `self.ntail` is 5 (because 5 bytes have been put into `self.tail`),
// and `needed` is therefore 3.
//
// - Scenario 1, `self.write_u8(0xFF)`: we have already zero-extended
// the input to 0x0000_0000_0000_00FF. We now left-shift it five
// bytes, giving 0x0000_FF00_0000_0000. We then bitwise-OR that value
// into `self.tail`, resulting in 0x0000_FFEE_DDCC_BBAA.
// (Zero-extension of the original input is critical in this scenario
// because we don't want the high two bytes of `self.tail` to be
// touched by the bitwise-OR.) `self.tail` is not yet full, so we
// return early, after updating `self.ntail` to 6.
//
// - Scenario 2, `self.write_u32(0xIIHH_GGFF)`: we have already
// zero-extended the input to 0x0000_0000_IIHH_GGFF. We now
// left-shift it five bytes, giving 0xHHGG_FF00_0000_0000. We then
// bitwise-OR that value into `self.tail`, resulting in
// 0xHHGG_FFEE_DDCC_BBAA. `self.tail` is now full, and we can use it
// to update `self.state`. (As mentioned above, this assumes a
// little-endian machine; on a big-endian machine we would have
// byte-swapped 0xIIHH_GGFF in the caller, giving 0xFFGG_HHII, and we
// would then end up bitwise-ORing 0xGGHH_II00_0000_0000 into
// `self.tail`).
//
self.tail |= x << (8 * self.ntail);
if size < needed {
self.ntail += size;
return;
}
// `self.tail` is full, process it.
self.state.v3 ^= self.tail;
Sip24Rounds::c_rounds(&mut self.state);
self.state.v0 ^= self.tail;
// Buffered tail is now flushed, process new input.
self.ntail = length - needed;
self.tail = unsafe { u8to64_le(msg, needed, self.ntail) };
}
#[inline(always)]
fn short_write_gen<T>(&mut self, x: T) {
let bytes =
unsafe { slice::from_raw_parts(&x as *const T as *const u8, mem::size_of::<T>()) };
self.short_write(bytes);
// Continuing scenario 2: we have one byte left over from the input. We
// set `self.ntail` to 1 and `self.tail` to `0x0000_0000_IIHH_GGFF >>
// 8*3`, which is 0x0000_0000_0000_00II. (Or on a big-endian machine
// the prior byte-swapping would leave us with 0x0000_0000_0000_00FF.)
//
// The `if` is needed to avoid shifting by 64 bits, which Rust
// complains about.
self.ntail = size - needed;
self.tail = if needed < 8 { x >> (8 * needed) } else { 0 };
}
#[inline]
@ -182,52 +215,52 @@ impl SipHasher128 {
impl Hasher for SipHasher128 {
#[inline]
fn write_u8(&mut self, i: u8) {
self.short_write_gen(i);
self.short_write(i, i as u64);
}
#[inline]
fn write_u16(&mut self, i: u16) {
self.short_write_gen(i);
self.short_write(i, i.to_le() as u64);
}
#[inline]
fn write_u32(&mut self, i: u32) {
self.short_write_gen(i);
self.short_write(i, i.to_le() as u64);
}
#[inline]
fn write_u64(&mut self, i: u64) {
self.short_write_gen(i);
self.short_write(i, i.to_le() as u64);
}
#[inline]
fn write_usize(&mut self, i: usize) {
self.short_write_gen(i);
self.short_write(i, i.to_le() as u64);
}
#[inline]
fn write_i8(&mut self, i: i8) {
self.short_write_gen(i);
self.short_write(i, i as u8 as u64);
}
#[inline]
fn write_i16(&mut self, i: i16) {
self.short_write_gen(i);
self.short_write(i, (i as u16).to_le() as u64);
}
#[inline]
fn write_i32(&mut self, i: i32) {
self.short_write_gen(i);
self.short_write(i, (i as u32).to_le() as u64);
}
#[inline]
fn write_i64(&mut self, i: i64) {
self.short_write_gen(i);
self.short_write(i, (i as u64).to_le() as u64);
}
#[inline]
fn write_isize(&mut self, i: isize) {
self.short_write_gen(i);
self.short_write(i, (i as usize).to_le() as u64);
}
#[inline]