Skip to content

Commit

Permalink
Auto merge of rust-lang#87794 - bonega:enum_niche_prefer_zero, r=nagisa
Browse files Browse the repository at this point in the history
Enum should prefer discriminant zero for niche

Given an enum with unassigned zero-discriminant, rust should prefer it for niche selection.
Zero as discriminant for `Option<Enum>` makes it possible for LLVM to optimize resulting asm.

- Eliminate branch when expected value coincides.
- Use smaller instruction `test eax, eax` instead of `cmp eax, ?`
- Possible interaction with zeroed memory?

Example:
```rust

pub enum Size {
    One = 1,
    Two = 2,
    Three = 3,
}

pub fn handle(x: Option<Size>) -> u8 {
    match x {
        None => {0}
        Some(size) => {size as u8}
    }
}
```
In this case discriminant zero is available as a niche.

Above example on nightly:
```asm
 mov     eax, edi
 cmp     al, 4
 jne     .LBB0_2
 xor     eax, eax
.LBB0_2:
 ret
```

PR:
```asm
 mov     eax, edi
 ret
```

I created this PR because I had a performance regression when I tried to use an enum to represent legal grapheme byte-length for utf8.

Using an enum instead of `NonZeroU8` [here](https://github.com/bonega/yore/blob/d683304f5dfe2e99f769e6ab8adf8d60a0d1d9b3/src/internal/decoder_incomplete.rs#L90)
resulted in a performance regression of about 5%.
I consider this to be a somewhat realistic benchmark.

Thanks to `@ogoffart` for pointing me in the right direction!

Edit: Updated description
  • Loading branch information
bors committed Sep 13, 2021
2 parents 9bb77da + 4d66fbc commit 9f85cd6
Show file tree
Hide file tree
Showing 3 changed files with 82 additions and 9 deletions.
52 changes: 43 additions & 9 deletions compiler/rustc_target/src/abi/mod.rs
Original file line number Diff line number Diff line change
Expand Up @@ -1099,19 +1099,53 @@ impl Niche {
assert!(size.bits() <= 128);
let max_value = size.unsigned_int_max();

if count > max_value {
let niche = v.end.wrapping_add(1)..v.start;
let available = niche.end.wrapping_sub(niche.start) & max_value;
if count > available {
return None;
}

// Compute the range of invalid values being reserved.
let start = v.end.wrapping_add(1) & max_value;
let end = v.end.wrapping_add(count) & max_value;

if v.contains(end) {
return None;
// Extend the range of valid values being reserved by moving either `v.start` or `v.end` bound.
// Given an eventual `Option<T>`, we try to maximize the chance for `None` to occupy the niche of zero.
// This is accomplished by prefering enums with 2 variants(`count==1`) and always taking the shortest path to niche zero.
// Having `None` in niche zero can enable some special optimizations.
//
// Bound selection criteria:
// 1. Select closest to zero given wrapping semantics.
// 2. Avoid moving past zero if possible.
//
// In practice this means that enums with `count > 1` are unlikely to claim niche zero, since they have to fit perfectly.
// If niche zero is already reserved, the selection of bounds are of little interest.
let move_start = |v: WrappingRange| {
let start = v.start.wrapping_sub(1) & max_value;
Some((start, Scalar { value, valid_range: v.with_start(start) }))
};
let move_end = |v: WrappingRange| {
let start = v.end.wrapping_add(1) & max_value;
let end = v.end.wrapping_add(count) & max_value;
Some((start, Scalar { value, valid_range: v.with_end(end) }))
};
let distance_end_zero = max_value - v.end;
if v.start > v.end {
// zero is unavailable because wrapping occurs
move_end(v)
} else if v.start <= distance_end_zero {
if count <= v.start {
move_start(v)
} else {
// moved past zero, use other bound
move_end(v)
}
} else {
let end = v.end.wrapping_add(count) & max_value;
let overshot_zero = (1..=v.end).contains(&end);
if overshot_zero {
// moved past zero, use other bound
move_start(v)
} else {
move_end(v)
}
}

Some((start, Scalar { value, valid_range: v.with_end(end) }))
}
}

Expand Down
25 changes: 25 additions & 0 deletions src/test/assembly/niche-prefer-zero.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,25 @@
// Check that niche selection prefers zero and that jumps are optimized away.
// See https://github.com/rust-lang/rust/pull/87794
// assembly-output: emit-asm
// only-x86
// compile-flags: -Copt-level=3

#![crate_type = "lib"]

#[repr(u8)]
pub enum Size {
One = 1,
Two = 2,
Three = 3,
}

#[no_mangle]
pub fn handle(x: Option<Size>) -> u8 {
match x {
None => 0,
Some(size) => size as u8,
}
}

// There should be no jumps in output
// CHECK-NOT: j
14 changes: 14 additions & 0 deletions src/test/ui/enum-discriminant/niche-prefer-zero.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,14 @@
// Check that niche selection prefers zero.
// See https://github.com/rust-lang/rust/pull/87794
// run-pass
#[repr(u8)]
pub enum Size {
One = 1,
Two = 2,
Three = 3,
}

fn main() {
// check that `None` is zero
assert_eq!(0, unsafe { std::mem::transmute::<Option<Size>, u8>(None) });
}

0 comments on commit 9f85cd6

Please sign in to comment.