OpenE2K
/
rust
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Rollup merge of #68787 - amosonn:patch-1, r=nagisa 

Optimize core::ptr::align_offset (part 1)

r? @nagisa
See #68616 for main discussion.
This commit is contained in:
Dylan DPC 2020-02-03 18:58:33 +01:00 committed by GitHub
parent 51c6c25575 22b263ae18
commit 370fd8c1b7
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 22 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
src/libcore/ptr/mod.rs
View File

@ -1081,9 +1081,8 @@ pub(crate) unsafe fn align_offset<T: Sized>(p: *const T, a: usize) -> usize {
// uses e.g., subtraction `mod n`. It is entirely fine to do them `mod
// usize::max_value()` instead, because we take the result `mod n` at the end
// anyway.
inverse = inverse.wrapping_mul(2usize.wrapping_sub(x.wrapping_mul(inverse)))
& (going_mod - 1);
if going_mod > m {
inverse = inverse.wrapping_mul(2usize.wrapping_sub(x.wrapping_mul(inverse)));
if going_mod >= m {
return inverse & (m - 1);
}
going_mod = going_mod.wrapping_mul(going_mod);
@ -1115,26 +1114,33 @@ pub(crate) unsafe fn align_offset<T: Sized>(p: *const T, a: usize) -> usize {
let gcdpow = intrinsics::cttz_nonzero(stride).min(intrinsics::cttz_nonzero(a));
let gcd = 1usize << gcdpow;
if p as usize & (gcd - 1) == 0 {
if p as usize & (gcd.wrapping_sub(1)) == 0 {
// This branch solves for the following linear congruence equation:
//
// $$ p + so ≡ 0 mod a $$
// ` p + so = 0 mod a `
//
// $p$ here is the pointer value, $s$ – stride of `T`, $o$ offset in `T`s, and $a$ – the
// `p` here is the pointer value, `s` - stride of `T`, `o` offset in `T`s, and `a` - the
// requested alignment.
//
// g = gcd(a, s)
// o = (a - (p mod a))/g * ((s/g)⁻¹ mod a)
// With `g = gcd(a, s)`, and the above asserting that `p` is also divisible by `g`, we can
// denote `a' = a/g`, `s' = s/g`, `p' = p/g`, then this becomes equivalent to:
//
// The first term is “the relative alignment of p to a”, the second term is “how does
// incrementing p by s bytes change the relative alignment of p”. Division by `g` is
// necessary to make this equation well formed if $a$ and $s$ are not co-prime.
// ` p' + s'o = 0 mod a' `
// ` o = (a' - (p' mod a')) * (s'^-1 mod a') `
//
// Furthermore, the result produced by this solution is not “minimal”, so it is necessary
// to take the result $o mod lcm(s, a)$. We can replace $lcm(s, a)$ with just a $a / g$.
let j = a.wrapping_sub(pmoda) >> gcdpow;
let k = smoda >> gcdpow;
return intrinsics::unchecked_rem(j.wrapping_mul(mod_inv(k, a)), a >> gcdpow);
// The first term is "the relative alignment of `p` to `a`" (divided by the `g`), the second
// term is "how does incrementing `p` by `s` bytes change the relative alignment of `p`" (again
// divided by `g`).
// Division by `g` is necessary to make the inverse well formed if `a` and `s` are not
// co-prime.
//
// Furthermore, the result produced by this solution is not "minimal", so it is necessary
// to take the result `o mod lcm(s, a)`. We can replace `lcm(s, a)` with just a `a'`.
let a2 = a >> gcdpow;
let a2minus1 = a2.wrapping_sub(1);
let s2 = smoda >> gcdpow;
let minusp2 = a2.wrapping_sub(pmoda >> gcdpow);
return (minusp2.wrapping_mul(mod_inv(s2, a2))) & a2minus1;
}
// Cannot be aligned at all.