Missed optimization opportunity in (z >= 0) && (x >= y + z) && (x >= y);

[krtab]

Hi!

I'm not too familiar LLVM. I was investigating this rustc issue rust-lang/rust#82801 and I realized that in the following Rust code, the last (x >= y) check is not optimized away:

#![feature(unchecked_math)]

#[no_mangle]
pub unsafe fn foo(x: u32, y : u32, z: u32) -> bool {
    (x >= y.unchecked_add(z)) && (x >= y)
}

I wrote what I believe is an equivalent C code, using the fact that overflow is UB on signed integers

bool foo(int x, int y, int z) {
    return (z >= 0) && (x >= y + z) && (x >= y);
}

but using clang trunk I get

foo:                                    # @foo
        test    edx, edx
        setns   cl
        add     edx, esi
        cmp     edx, edi
        setle   dl
        cmp     edi, esi
        setge   al
        and     al, cl
        and     al, dl
        ret

whereas GCC properly removes the last comparison

foo:
        xor     eax, eax
        test    edx, edx
        js      .L1
        add     edx, esi
        cmp     edx, edi
        setle   al
.L1:
        ret

edit:

Weirdly enough, the code below is well optimized by clang 17 but not by goldbot's clang trunk version, so there is a possible regression.

bool foo(int x, int y, int z) {
    if (z >= 0) if (x >= y + z) return (x >= y);
    return (z >= 0) && (x >= y + z) && (x >= y);
}

[dtcxzyw]

https://alive2.llvm.org/ce/z/v5GPts

[pinskia]

I noticed GCC does not optimize the case where s/&&/& is done, filed https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112386 for that.

[krtab]

Thanks @dtcxzyw !