[llvm-bugs] [Bug 51075] New: Failure to make use of dereference range to improve vectorization

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=51075

            Bug ID: 51075
           Summary: Failure to make use of dereference range to improve
                    vectorization
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: llvm-dev at redking.me.uk
                CC: a.bataev at hotmail.com, llvm-bugs at lists.llvm.org,
                    spatel+llvm at rotateright.com

https://simd.godbolt.org/z/PTsGGYzeW

As mentioned on [Bug #50920] we often have cases where float3 types are
embedded with a float[4]/float4 payload. Typically to assist is some custom
downcoding for particularly hot code, but where scalar code is used otherwise.

struct Vector3 {
    float m_floats[4];
};

Vectorization of this float dotproduct3 scalar code results in a mixture of
float2 and float types:

float Dot3(struct Vector3 &a, struct Vector3 &b) {
    float dot = 0;
    dot += a.m_floats[0] * b.m_floats[0];
    dot += a.m_floats[1] * b.m_floats[1];
    dot += a.m_floats[2] * b.m_floats[2];
    return dot;
}

define float @Dot3(%struct.Vector3* nocapture nonnull readonly align 4
dereferenceable(16) %0, %struct.Vector3* nocapture nonnull readonly align 4
dereferenceable(16) %1) {
  %3 = getelementptr inbounds %struct.Vector3, %struct.Vector3* %0, i64 0, i32
0, i64 0
  %4 = load float, float* %3, align 4, !tbaa !3
  %5 = getelementptr inbounds %struct.Vector3, %struct.Vector3* %1, i64 0, i32
0, i64 0
  %6 = load float, float* %5, align 4, !tbaa !3
  %7 = fmul float %4, %6
  %8 = fadd float %7, 0.000000e+00
  %9 = getelementptr inbounds %struct.Vector3, %struct.Vector3* %0, i64 0, i32
0, i64 1
  %10 = getelementptr inbounds %struct.Vector3, %struct.Vector3* %1, i64 0, i32
0, i64 1
  %11 = bitcast float* %9 to <2 x float>*
  %12 = load <2 x float>, <2 x float>* %11, align 4, !tbaa !3
  %13 = bitcast float* %10 to <2 x float>*
  %14 = load <2 x float>, <2 x float>* %13, align 4, !tbaa !3
  %15 = fmul <2 x float> %12, %14
  %16 = extractelement <2 x float> %15, i32 0
  %17 = fadd float %8, %16
  %18 = extractelement <2 x float> %15, i32 1
  %19 = fadd float %17, %18
  ret float %19
}

But we should make use of the fact that the sources are both dereferencable as
float4, and then either extract the float2/float inside or maybe treat this as
a <float x 3>:

float Dot3v(struct Vector3 &a, struct Vector3 &b) {
    __m128 aa = _mm_loadu_ps(a.m_floats);
    __m128 bb = _mm_loadu_ps(b.m_floats);
    __m128 rr = _mm_mul_ps(aa, bb);
    __m128 dot = rr;
    dot = _mm_add_ss(dot, _mm_shuffle_ps(rr, rr, _MM_SHUFFLE(1,1,1,1)));
    dot = _mm_add_ss(dot, _mm_shuffle_ps(rr, rr, _MM_SHUFFLE(2,2,2,2)));
    return _mm_cvtss_f32(dot);
}

define float @Dot3v(%struct.Vector3* nocapture nonnull readonly align 4
dereferenceable(16) %0, %struct.Vector3* nocapture nonnull readonly align 4
dereferenceable(16) %1) {
  %3 = bitcast %struct.Vector3* %0 to <4 x float>*
  %4 = load <4 x float>, <4 x float>* %3, align 4, !tbaa !3
  %5 = bitcast %struct.Vector3* %1 to <4 x float>*
  %6 = load <4 x float>, <4 x float>* %5, align 4, !tbaa !3
  %7 = fmul <4 x float> %4, %6
  %8 = shufflevector <4 x float> %7, <4 x float> poison, <4 x i32> <i32 1, i32
undef, i32 undef, i32 undef>
  %9 = fadd <4 x float> %7, %8
  %10 = shufflevector <4 x float> %7, <4 x float> poison, <4 x i32> <i32 2, i32
undef, i32 undef, i32 undef>
  %11 = fadd <4 x float> %10, %9
  %12 = extractelement <4 x float> %11, i64 0
  ret float %12
}

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20210713/9c593cea/attachment.html>