| Issue |
184245
|
| Summary |
Miscompilation of `VMINMAXS{H,S,D}` intrinsics
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
sayantn
|
Intrinsics such as `_mm_mask_minmax_ss` should copy the upper elements from the first argument (`a`), but the LLVM implementation seems to copy it from the fallback argument (`src`).
[Intel Advamced Vector Extensions 10.2 Architecture Specification, Revision 6.0](https://www.intel.com/content/www/us/en/content-details/913918/intel-advanced-vector-extensions-10-2-intel-avx10-2-architecture-specification.html), Page 190
> Bits: 127:16/32/64 are copied to the destination from the respective elements of the first operand
As a concrete example, consider
`experiment.c`
```C
#include<immintrin.h>
#include<stdio.h>
int main() {
__m128 a = _mm_setr_ps(1.0, 2.0, 3.0, 4.0);
__m128 b = _mm_setr_ps(5.0, 6.0, 7.0, 8.0);
__m128 src = "" 30.0, 40.0, 50.0);
__m128 r = _mm_mask_minmax_ss(src, 1, a, b, 0);
float array[4];
_mm_storeu_ps(array, r);
for (int i =0; i < 4; ++i) {
printf("%f\n", array[i]);
}
}
```
<img width="843" height="342" alt="Image" src="" />
The versions are as follows
`clang --version`
```
ClangBuiltLinux clang version 22.1.0 (https://github.com/llvm/llvm-project.git 4434dabb69916856b824f68a64b029c67175e532)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /home/schak/Documents/llvm-22.1.0-x86_64/bin
```
`gcc --version`
```
gcc (GCC) 15.2.1 20260209
Copyright (C) 2025 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
```
`sde64 --version`
```
Intel(R) Software Development Emulator. Version: 10.5.0 external (0)
Copyright (C) 2008-2026, Intel Corporation. All rights reserved.
```
The equivalent(?) program in Rust behaves a bit differently (Rust doesn't yet have AVX10.2 intrinsics, in fact I am currently drafting a PR to add them, that's how I came across this)
`main.rs`
```rust
#![feature(link_llvm_intrinsics, abi_unadjusted, simd_ffi, avx10_target_feature)]
#![allow(internal_features)]
use std::arch::x86_64::*;
#[allow(improper_ctypes)]
unsafe extern "unadjusted" {
#[link_name = "llvm.x86.avx10.mask.vminmaxss.round"]
fn vminmaxss(a: __m128, b: __m128, imm8: i32, src: __m128, k: __mmask8, sae: i32) -> __m128;
}
#[inline]
#[target_feature(enable = "avx10.2")]
pub fn _mm_mask_minmax_ss<const IMM8: i32>(
src: __m128,
k: __mmask8,
a: __m128,
b: __m128,
) -> __m128 {
unsafe { vminmaxss(a, b, IMM8, src, k as u8, _MM_FROUND_CUR_DIRECTION) }
}
fn main() {
unsafe {
let a = _mm_setr_ps(1.0, 2.0, 3.0, 4.0);
let b = _mm_setr_ps(5.0, 6.0, 7.0, 8.0);
let src = "" 30.0, 40.0, 50.0);
let r = _mm_mask_minmax_ss::<0>(src, 1, a, b);
let mut array = [0.0f32; 4];
_mm_storeu_ps(array.as_mut_ptr(), r);
for elem in array {
println!("{elem}")
}
}
}
```
<img width="900" height="170" alt="Image" src="" />
`rustc --version --verbose`
```
rustc 1.96.0-nightly (80381278a 2026-03-01)
binary: rustc
commit-hash: 80381278a08582356c13b0f52af92d27c567c230
commit-date: 2026-03-01
host: x86_64-unknown-linux-gnu
release: 1.96.0-nightly
LLVM version: 22.1.0
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
