Missed optimization with core::cmp::{min,max} {<,>} n and the range of one the values is known

[paolobarbolini]

I tried this code:

#![feature(core_intrinsics)]

pub unsafe fn foo1(a: usize, b: usize) {
    std::intrinsics::assume(a < 6);

    assert!(a.min(b) < 6);
}

I expected to see this happen: the assert is optimized away. By proving a < 6, even without knowing b, we can assume a.min(b) < 6.

Instead, this happened: it doesn't optimize, but it does in the following conditions:

#![feature(core_intrinsics)]

pub unsafe fn foo2(a: usize, b: usize) {
    std::intrinsics::assume(a < 6);

    assert!(a < 6 || b < 6);

    assert!((if a < b { a } else { b }) < 6);
    assert!((if a <= b { a } else { b }) < 6);
}

// or...

pub unsafe fn foo3(a: usize, b: usize) {
    std::intrinsics::assume(a == 3);

    assert!(a.min(b) < 6);
}


// or...

pub unsafe fn foo4(a: usize, b: usize) {
    std::intrinsics::assume(a < 6);
    std::intrinsics::assume(b < 6);

    assert!(a.min(b) < 6);
}

Why this matters

Prevents the ptr_rotate implementation from optimizing here

rust/library/core/src/slice/rotate.rs

Line 161 in 7be8693

} else if cmp::min(left, right) <= mem::size_of::<BufType>() / mem::size_of::<T>() {

which would reduce the amount of assembly generated by #89714

Meta

rustc --version --verbose:

rustc 1.60.0-nightly (ec4bcaac4 2022-01-15)
binary: rustc
commit-hash: ec4bcaac450279b029f3480b8b8f1b82ab36a5eb
commit-date: 2022-01-15
host: x86_64-unknown-linux-gnu
release: 1.60.0-nightly
LLVM version: 13.0.0

[nikic]

Godbolt: https://rust.godbolt.org/z/78sGo6na7 Still doesn't optimize on LLVM head: https://llvm.godbolt.org/z/nTPM51W7K

It looks like LVI currently only handles the case where both sides of the min/max have a known range: https://github.com/llvm/llvm-project/blob/30a4020a7db866990199f523a60ad474d7ec3efa/llvm/lib/Analysis/LazyValueInfo.cpp#L810 Should be easy to fix at least.

[nikic]

Should be fixed by llvm/llvm-project@202d590, llvm/llvm-project@d15823e, llvm/llvm-project@9e68557.

[paolobarbolini]

I was able to confirm that #93577 fixed it
https://rust.godbolt.org/z/o6bqsKEjs

Closing