32x performance regression for AVX2 intrinsics in Rust v1.87

[Shnatsel]

Reproduction steps

git clone https://github.com/mcountryman/simd-adler32
cd simd-adler32
cargo +1.86.0 bench
cargo +1.87.0 bench

On AVX2-capable CPUs, on Rust 1.86 you can see the AVX2 codepath is performing well, at or above the SSE3 level.

On Rust 1.87 performance collapses completely, with the AVX2 implementation being the slowest, behind even the scalar implementation.

SSE2, SSE3 and AVX-512 are not affected.

The simd-adler32 crate uses runtime feature detection and explicit AVX2 intrinsics from std::arch.

Version it worked on

1.86

Version with regression

1.87

rustc --version --verbose:

rustc 1.87.0 (17067e9ac 2025-05-09)
binary: rustc
commit-hash: 17067e9ac6d7ecb70e50f92c1944e545188d2359
commit-date: 2025-05-09
host: x86_64-unknown-linux-gnu
release: 1.87.0
LLVM version: 20.1.1

[workingjubilee]

Does this repro on nightly?

[Shnatsel]

It is not an issue on the latest nightly:

rustc 1.89.0-nightly (586ad391f 2025-06-15)
binary: rustc
commit-hash: 586ad391f5ee4519acc7cae340e34673bae762b1
commit-date: 2025-06-15
host: x86_64-unknown-linux-gnu
release: 1.89.0-nightly
LLVM version: 20.1.5

[workingjubilee]

beta:

variants/avx2-10b       time:   [4.3867 ns 4.4002 ns 4.4148 ns]                               
                        thrpt:  [2.1095 GiB/s 2.1166 GiB/s 2.1230 GiB/s]
                 change:
                        time:   [-27.344% -27.114% -26.882%] (p = 0.00 < 0.05)
                        thrpt:  [+36.766% +37.200% +37.635%]
                        Performance has improved.
Found 5 outliers among 100 measurements (5.00%)
  5 (5.00%) high mild
variants/avx2-10k       time:   [144.94 ns 145.11 ns 145.27 ns]                              
                        thrpt:  [64.109 GiB/s 64.181 GiB/s 64.256 GiB/s]
                 change:
                        time:   [-98.792% -98.789% -98.787%] (p = 0.00 < 0.05)
                        thrpt:  [+8140.7% +8160.3% +8179.8%]
                        Performance has improved.
Found 4 outliers among 100 measurements (4.00%)
  4 (4.00%) low mild
variants/avx2-100k      time:   [1.4382 µs 1.4408 µs 1.4435 µs]                                
                        thrpt:  [64.517 GiB/s 64.639 GiB/s 64.756 GiB/s]
                 change:
                        time:   [-98.791% -98.786% -98.782%] (p = 0.00 < 0.05)
                        thrpt:  [+8110.1% +8139.9% +8168.7%]
                        Performance has improved.
Found 1 outliers among 100 measurements (1.00%)
  1 (1.00%) high mild

sorry for the inconvenience, please wait ~10 days for this to be fixed!

[nikic]

Do we have any idea what caused/fixed this? I initially thought it might be #139029, but that shouldn't affect the avx2 code working on __m256i, right?

[tgross35]

I was thinking the same thing at first - but you are correct, #135408 only touched vector types up to 128 bits.