Buggy optimization of vfmaddcsh intrinsics

[sayantn]

The llvm.x86.avx512fp16.maskz.vfmadd.csh intrinsic (and due to that, _mm_maskz_fmadd_sch) is being incorrectly optimized. This code snippet

#include<immintrin.h>
#include<stdio.h>

int main() {
    __m128h a, b, c, r;
    _Float16 array[8];

    a = _mm_setr_ph(0.0, 1.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0);
    b = _mm_setr_ph(0.0, 2.0, 16.0, 17.0, 18.0, 19.0, 20.0, 21.0);
    c = _mm_setr_ph(0.0, 3.0, 22.0, 23.0, 24.0, 25.0, 26.0, 27.0);

    r = _mm_maskz_fmadd_sch(0, a, b, c);
    _mm_storeu_ph(array, r);

    for (int i = 0; i < 8; i++){
        printf("%f\n", (float) array[i]);
    }

    return 0;
}

In clang, the unoptimized and optimized output is different. The unoptimized output is the correct one according to Intel. gcc gives the correct output in both.

System specification:

• mingw-w64-x86_64-gcc 14.2.0
• mingw-w64-x86_64-clang 18.1.8
• Intel Software Development Emulator v9.44.0

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Sayantan Chakraborty (sayantn)

The `llvm.x86.avx512fp16.maskz.vfmadd.csh` intrinsic (and due to that, `_mm_maskz_fmadd_sch`) is being incorrectly optimized. This code snippet

#include<immintrin.h>
#include<stdio.h>

int main() {
    __m128h a, b, c, r;
    _Float16 array[8];

    a = _mm_setr_ph(0.0, 1.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0);
    b = _mm_setr_ph(0.0, 2.0, 16.0, 17.0, 18.0, 19.0, 20.0, 21.0);
    c = _mm_setr_ph(0.0, 3.0, 22.0, 23.0, 24.0, 25.0, 26.0, 27.0);

    r = _mm_maskz_fmadd_sch(0, a, b, c);
    _mm_storeu_ph(array, r);

    for (int i = 0; i < 8; i++){
        printf("%f\n", (float) array[i]);
    }

    return 0;
}

In clang, the unoptimized and optimized output is different. The unoptimized output is the correct one according to Intel. gcc gives the correct output in both.

System specification:

• mingw-w64-x86_64-gcc 14.1.0-3
• mingw-w64-x86_64-clang 18.1.8-1
• Intel Software Development Emulator v9.33.0

[RKSimon]

CC @phoebewang @KanRobert