Bounds check is not eliminated

[SeeSpring]

I tried this code:
Godbolt

#![feature(portable_simd)]
use std::simd::*;
pub fn increment(idx: Simd<u32, 8>, arr: &mut [u32]) {
    assert!(
        (arr.len() >= u32::MAX as usize)
            || idx.simd_lt(Simd::splat(arr.len() as u32)).all()
    );
    idx.as_array().iter().for_each(|i| arr[*i as usize] += 1);
}

I expected to see this happen: Bounds checks are eliminated and there is the vpcmpltd instruction

Instead, this happened: Bounds checks are not eliminated and a vpmaxud and vpcmpeqd pair is used

Meta

RUSTC_BOOTSTRAP=1 rustc --version --verbose:

rustc 1.70.0 (90c541806 2023-05-31)
binary: rustc
commit-hash: 90c541806f23a127002de5b4038be731ba1458ca
commit-date: 2023-05-31
host: x86_64-unknown-linux-gnu
release: 1.70.0
LLVM version: 16.0.2

[SeeSpring]

Also vmovmskps and test should probably just be vtestps

[calebzulawski]

It appears that on latest Rust, the codegen is as expected, using vcmp + test: godbolt

[SeeSpring]

There are still the bounds checks in the disassembly

.LBB0_2:
        mov     ecx, dword ptr [rax]
        mov     edi, ecx
        cmp     rdi, rdx
        jae     .LBB0_12
        inc     dword ptr [rsi + 4*rdi]
        mov     edi, dword ptr [rax + 4]
        cmp     rdi, rdx
        jae     .LBB0_12
        inc     dword ptr [rsi + 4*rdi]
        mov     edi, dword ptr [rax + 8]
        cmp     rdi, rdx
        jae     .LBB0_12
        inc     dword ptr [rsi + 4*rdi]
        mov     edi, dword ptr [rax + 12]
        cmp     rdi, rdx
        jae     .LBB0_12
        inc     dword ptr [rsi + 4*rdi]
        mov     edi, dword ptr [rax + 16]
        cmp     rdi, rdx
        jae     .LBB0_12
        inc     dword ptr [rsi + 4*rdi]
        mov     edi, dword ptr [rax + 20]
        cmp     rdi, rdx
        jae     .LBB0_12
        inc     dword ptr [rsi + 4*rdi]
        mov     edi, dword ptr [rax + 24]
        cmp     rdi, rdx
        jae     .LBB0_12
        inc     dword ptr [rsi + 4*rdi]
        mov     edi, dword ptr [rax + 28]
        cmp     rdi, rdx
        jae     .LBB0_12
        inc     dword ptr [rsi + 4*rdi]
        pop     rax
        vzeroupper
        ret

When using unsafe { idx.as_array().iter().for_each(|i| *arr.get_unchecked_mut(*i as usize) += 1); }

.LBB0_3:
        vmovd   eax, xmm0
        inc     dword ptr [rsi + 4*rax]
        vpextrd eax, xmm0, 1
        inc     dword ptr [rsi + 4*rax]
        vpextrd eax, xmm0, 2
        inc     dword ptr [rsi + 4*rax]
        vpextrd eax, xmm0, 3
        inc     dword ptr [rsi + 4*rax]
        vextracti128    xmm0, ymm0, 1
        vmovd   eax, xmm0
        inc     dword ptr [rsi + 4*rax]
        vpextrd eax, xmm0, 1
        inc     dword ptr [rsi + 4*rax]
        vpextrd eax, xmm0, 2
        inc     dword ptr [rsi + 4*rax]
        vpextrd eax, xmm0, 3
        inc     dword ptr [rsi + 4*rax]
        pop     rax
        vzeroupper
        ret

[calebzulawski]

I don't think that's particularly SIMD-related, that's [u32]'s Index trait. I understand technically that assert proves removing the check is allowed, but that's a relatively complex relationship that I'm not sure LLVM can handle.

[calebzulawski]

As a comparison, the trivial scalar implementation doesn't remove the bound checks either: godbolt