[Optimization] Significantly different assembly output for logically equivalent match patterns with string comparisons

[ZhonghaoPan-nju]

I tried these codes:
https://godbolt.org/z/c3Ynrcb7j

#[no_mangle]
fn longest<'a>(s1: &'a str, s2: &'a str) -> &'a str {
    match s1.len().cmp(&s2.len()) {
        std::cmp::Ordering::Greater|std::cmp::Ordering::Less if s1 != "\"" || s2 != "" => { s1 }
        std::cmp::Ordering::Greater => s1,
        std::cmp::Ordering::Less => s2,
        std::cmp::Ordering::Equal => s1,
    }
}

and:

#[no_mangle]
fn longest<'a>(s1: &'a str, s2: &'a str) -> &'a str {
    let temp_bridge_9 = s1;
    match temp_bridge_9.len().cmp(&s2.len()) {
        std::cmp::Ordering::Greater | std::cmp::Ordering::Less if !(s1 == "\"") || !(s2 == "") => {
            s1
        }
        std::cmp::Ordering::Greater => s1,
        std::cmp::Ordering::Less => s2,
        std::cmp::Ordering::Equal => s1,
    }
}

I expected to see this happen:

longest:
        mov     rdx, rsi
        mov     rax, rdi
        ret

Instead, this happened:

longest:
        mov     r8, rdx
        mov     rdx, rsi
        mov     rax, rdi
        cmp     rsi, rcx
        seta    sil
        sbb     sil, 0
        je      .LBB0_10
        movzx   esi, sil
        cmp     esi, 1
        jne     .LBB0_6
        cmp     rdx, 1
        jne     .LBB0_10
        test    rcx, rcx
        jne     .LBB0_10
        cmp     byte ptr [rax], 34
        jne     .LBB0_10
        mov     edx, 1
.LBB0_10:
        ret
.LBB0_6:
        cmp     rdx, 1
        jne     .LBB0_10
        test    rcx, rcx
        jne     .LBB0_10
        cmp     byte ptr [rax], 34
        jne     .LBB0_10
        xor     edx, edx
        mov     rax, r8
        ret

There are notable discrepancies in the assembly output between the two functionally equivalent implementations of longest. I suspect that the variant logic in the condition s1 != "\"" || s2 != "" may be influencing the compiler's optimization behavior.

I’d greatly appreciate it if you could review these cases.
Thank you for your time and consideration!

Meta

rustc 1.85.0-nightly (d117b7f21 2024-12-31)
binary: rustc
commit-hash: d117b7f211835282b3b177dc64245fff0327c04c
commit-date: 2024-12-31
host: x86_64-unknown-linux-gnu
release: 1.85.0-nightly
LLVM version: 19.1.6

[saethlin]

Alive2 says the the optimization is valid: https://alive2.llvm.org/ce/z/4amEw6

[ZhonghaoPan-nju]

@saethlin If alive2 confirms it's valid, would it make sense to implement this optimization in LLVM/Rust?

[saethlin]

When alive2 says a transformation is valid between the two optimized LLVM IR outputs, it means that the improvement can be implemented in LLVM.

If alive2 doesn't say that the transformation is valid, sometimes the two programs just behave differently, and sometimes we are not giving LLVM all the information that makes the optimization valid.

I've added A-LLVM and I-slow because there are some LLVM contributors that watch our issue tracker for those labels.

[ZhonghaoPan-nju]

@saethlin I see. Thank you for your explanation. I truly hope this issue can contribute to rustc and LLVM.