clang/llvm fails to auto vector vector compare and select

[wzssyqa]

float *f(float *a, float *b, float *c) {
        for (int i=0; i<4; i++) {
                c[i] = a[i]>b[i]?a[i]:b[i];
        }
        return c;
}

./bin/clang -O3 --target=x86_64-linux-gnu -mavx2 -S n.c -emit-llvm -ffast-math emits IR like

define dso_local noundef ptr @f(ptr nocapture noundef readonly %a, ptr nocapture noundef readonly %b, ptr noundef returned writeonly %c) local_unnamed_addr #0 {
entry:
  %0 = load float, ptr %a, align 4, !tbaa !5
  %1 = load float, ptr %b, align 4, !tbaa !5
  %. = tail call fast float @llvm.maxnum.f32(float %0, float %1)
  store float %., ptr %c, align 4, !tbaa !5
  %arrayidx.1 = getelementptr inbounds i8, ptr %a, i64 4
  %2 = load float, ptr %arrayidx.1, align 4, !tbaa !5
  %arrayidx2.1 = getelementptr inbounds i8, ptr %b, i64 4
  %3 = load float, ptr %arrayidx2.1, align 4, !tbaa !5
  %..1 = tail call fast float @llvm.maxnum.f32(float %2, float %3)
  %arrayidx9.1 = getelementptr inbounds i8, ptr %c, i64 4
  store float %..1, ptr %arrayidx9.1, align 4, !tbaa !5
  %arrayidx.2 = getelementptr inbounds i8, ptr %a, i64 8
  %4 = load float, ptr %arrayidx.2, align 4, !tbaa !5
  %arrayidx2.2 = getelementptr inbounds i8, ptr %b, i64 8
  %5 = load float, ptr %arrayidx2.2, align 4, !tbaa !5
  %..2 = tail call fast float @llvm.maxnum.f32(float %4, float %5)
  %arrayidx9.2 = getelementptr inbounds i8, ptr %c, i64 8
  store float %..2, ptr %arrayidx9.2, align 4, !tbaa !5
  %arrayidx.3 = getelementptr inbounds i8, ptr %a, i64 12
  %6 = load float, ptr %arrayidx.3, align 4, !tbaa !5
  %arrayidx2.3 = getelementptr inbounds i8, ptr %b, i64 12
  %7 = load float, ptr %arrayidx2.3, align 4, !tbaa !5
  %..3 = tail call fast float @llvm.maxnum.f32(float %6, float %7)
  %arrayidx9.3 = getelementptr inbounds i8, ptr %c, i64 12
  store float %..3, ptr %arrayidx9.3, align 4, !tbaa !5
  ret ptr %c
}

The pair of [‘llvm.vp.fcmp.*’](https://llvm.org/docs/LangRef.html#id2420) and [‘llvm.vp.select.*’](https://llvm.org/docs/LangRef.html#id2349) may be better.

[RKSimon]

aliasing issue, the vectorizer can't confirm that c doesn't alias with a/b - this works fine:

float *f(float *a, float *b, float * __restrict c) {
        for (int i=0; i<4; i++) {
                c[i] = a[i]>b[i]?a[i]:b[i];
        }
        return c;
}

[RKSimon]

@wzssyqa OK to close this?