[AArch64] vbfdotq_laneq_f32 not generating indexed bfdot

[dsharlet]

Reproducer: https://godbolt.org/z/vdzs9nYWb

It's short enough to reproduce here, input:

#include <arm_neon.h>

float32x4_t dot_a_few(bfloat16x8_t a, bfloat16x8_t b) {
    float32x4_t result = vdupq_n_f32(0.0f);
    result = vbfdotq_laneq_f32(result, a, b, 0);
    result = vbfdotq_laneq_f32(result, a, b, 1);
    result = vbfdotq_laneq_f32(result, a, b, 2);
    result = vbfdotq_laneq_f32(result, a, b, 3);
    return result;
}

float32x4_t dot_a_few(bfloat16x4_t a, bfloat16x8_t b) {
    float32x4_t result = vdupq_n_f32(0.0f);
    result = vbfdotq_lane_f32(result, b, a, 0);
    result = vbfdotq_lane_f32(result, b, a, 1);
    return result;
}

Output:

dot_a_few(__Bfloat16x8_t, __Bfloat16x8_t):
        movi    v2.2d, #0000000000000000
        dup     v3.4s, v1.s[0]
        bfdot   v2.4s, v0.8h, v3.8h
        dup     v3.4s, v1.s[1]
        bfdot   v2.4s, v0.8h, v3.8h
        dup     v3.4s, v1.s[2]
        dup     v1.4s, v1.s[3]
        bfdot   v2.4s, v0.8h, v3.8h
        bfdot   v2.4s, v0.8h, v1.8h
        mov     v0.16b, v2.16b
        ret

dot_a_few(__Bfloat16x4_t, __Bfloat16x8_t):
        movi    v2.2d, #0000000000000000
        bfdot   v2.4s, v1.8h, v0.2h[0]
        bfdot   v2.4s, v1.8h, v0.2h[1]
        mov     v0.16b, v2.16b
        ret

Note that the vbfdotq_lane_f32 works, and generates an indexed bfdot, but vbfdotq_laneq_f32 does not, it's generating explicit dup instructions.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Dillon (dsharlet)

Reproducer: https://godbolt.org/z/vdzs9nYWb

It's short enough to reproduce here, input:

#include <arm_neon.h>

float32x4_t dot_a_few(bfloat16x8_t a, bfloat16x8_t b) {
    float32x4_t result = vdupq_n_f32(0.0f);
    result = vbfdotq_laneq_f32(result, a, b, 0);
    result = vbfdotq_laneq_f32(result, a, b, 1);
    result = vbfdotq_laneq_f32(result, a, b, 2);
    result = vbfdotq_laneq_f32(result, a, b, 3);
    return result;
}

float32x4_t dot_a_few(bfloat16x4_t a, bfloat16x8_t b) {
    float32x4_t result = vdupq_n_f32(0.0f);
    result = vbfdotq_lane_f32(result, b, a, 0);
    result = vbfdotq_lane_f32(result, b, a, 1);
    return result;
}

Output:

dot_a_few(__Bfloat16x8_t, __Bfloat16x8_t):
        movi    v2.2d, #<!-- -->0000000000000000
        dup     v3.4s, v1.s[0]
        bfdot   v2.4s, v0.8h, v3.8h
        dup     v3.4s, v1.s[1]
        bfdot   v2.4s, v0.8h, v3.8h
        dup     v3.4s, v1.s[2]
        dup     v1.4s, v1.s[3]
        bfdot   v2.4s, v0.8h, v3.8h
        bfdot   v2.4s, v0.8h, v1.8h
        mov     v0.16b, v2.16b
        ret

dot_a_few(__Bfloat16x4_t, __Bfloat16x8_t):
        movi    v2.2d, #<!-- -->0000000000000000
        bfdot   v2.4s, v1.8h, v0.2h[0]
        bfdot   v2.4s, v1.8h, v0.2h[1]
        mov     v0.16b, v2.16b
        ret

Note that the vbfdotq_lane_f32 works, and generates an indexed bfdot, but vbfdotq_laneq_f32 does not, it's generating explicit dup instructions.

[dsharlet]

Also, I just noticed, that while this reproducer is showing an issue, that in my real code, neither of these is generating an indexed bfdot. So maybe the issue here isn't necessarily that one or the other is working, but that the optimization is fragile?