`-C linker-plugin-lto` prevents vectorization of `mem::swap` at `-C opt-level=3`

[DaniPopes]

I tried this code (godbolt):

pub type T = [u64; 4];

#[no_mangle]
pub fn swap_32(a: &mut T, b: &mut T) {
    std::mem::swap(a, b);
    unsafe {
        // std::ptr::swap(a, b);
        // std::ptr::swap_nonoverlapping(a, b, 1);
    }
}

I expected to see this happen: code optimizes to a few <_ x i64> (target-cpu-dependent) loads and stores on -C opt-level=3.

Instead, this happened: the code is not vectorized when -C linker-plugin-lto is also passed to the compiler (this is passed together with -Clto=... but that has no effect here)

Note that this only happens with mem::swap/ptr::swap_nonoverlapping, not ptr::swap.

Meta

rustc --version --verbose:

rustc 1.79.0-nightly (f9b161492 2024-04-19)
binary: rustc
commit-hash: f9b16149208c8a8a349c32813312716f6603eb6f
commit-date: 2024-04-19
host: x86_64-unknown-linux-gnu
release: 1.79.0-nightly
LLVM version: 18.1.4

This happens since at least 1.61, but I could not find a way to bisect this, so I opted not to mark this a regression.

@rustbot modify labels: +A-LLVM +A-LTO +I-slow

[DaniPopes]

searched toolchains nightly-2020-12-25 through nightly-2024-04-21
Regression in nightly-2022-02-26
get_commits_between returning commits, len: 11
  commit[0] 2022-02-24: Auto merge of #94131 - Mark-Simulacrum:fmt-string, r=oli-obk
  commit[1] 2022-02-24: Auto merge of #94333 - Dylan-DPC:rollup-7yxtywp, r=Dylan-DPC
  commit[2] 2022-02-25: Auto merge of #93368 - eddyb:diagbld-guarantee, r=estebank
  commit[3] 2022-02-25: Auto merge of #93878 - Aaron1011:newtype-macro, r=cjgillot
  commit[4] 2022-02-25: Auto merge of #94130 - erikdesjardins:partially, r=nikic
  commit[5] 2022-02-25: Auto merge of #94350 - matthiaskrgr:rollup-eesfiyr, r=matthiaskrgr
  commit[6] 2022-02-25: Auto merge of #93644 - michaelwoerister:simpler-debuginfo-typemap, r=wesleywiser
  commit[7] 2022-02-25: Auto merge of #94357 - matthiaskrgr:rollup-xrjaof3, r=matthiaskrgr
  commit[8] 2022-02-25: Auto merge of #94279 - tmiasko:write-print, r=Mark-Simulacrum
  commit[9] 2022-02-25: Auto merge of #94290 - Mark-Simulacrum:bump-bootstrap, r=pietroalbini
  commit[10] 2022-02-25: Auto merge of #94369 - matthiaskrgr:rollup-qtripm2, r=matthiaskrgr
ERROR: no CI builds available between 4b043faba34ccc053a4d0110634c323f6c03765e and d3ad51b48f83329fac0cd8a9f1253f3146613c1c within last 167 days

4b043fa...d3ad51b

Probably #94212 cc @scottmcm

[scottmcm]

We used to basically be forcing this, but stopped doing that in #123185

I know nothing about what linker-plugin-lto does or why it would be relevant.

But yeah, in assembly it swaps via ymms in v3 https://rust.godbolt.org/z/Mqc64zYxW without the flag, but does individual qwords with it 🤷

[nikic]

I'm not sure whether -C linker-plugin-lto is compatible with --emit=llvm-ir (or --emit=asm). I'd guess it prints the pre-LTO IR in that case?

[DaniPopes]

My workflow is getting the --release assembly with cargo-show-asm. LTO is enabled in the release profile overrides in the manifest.

It's unfortunate that this does not return the fully optimized assembly, but it also makes sense since linking doesn't actually happen. I thought this issue initially was a regression because couldn't find any info on this.

I confirmed that it gets vectorized in the final assembly with objdump + grep in a new crate, so I'll close this issue.

[scottmcm]

Glad to hear it's working! (and that I didn't break everything *phew*)

TBH, this has been confusing me too as I work on stuff. Thus tests like https://github.com/rust-lang/rust/blob/master/tests/assembly/x86_64-typed-swap.rs as it seems that looking in LLVM IR keeps not telling the full story :(