[PATCH] D87391: [Intrinsics] define semantics for experimental fmax/fmin vector reductions

[Sanjay Patel via Phabricator via llvm-commits]

spatel added inline comments.


================
Comment at: llvm/test/CodeGen/Thumb2/mve-vecreduce-fminmax.ll:247
+; CHECK-FP-NEXT:    vminnm.f32 s0, s0, s1
+; CHECK-FP-NEXT:    vminnm.f32 s0, s0, s4
 ; CHECK-FP-NEXT:    bx lr
----------------
nikic wrote:
> This lowering looks incorrect for the case where both elements are NaN. We'll fold to +INF then. We probably have an expansion that assumes +INF is a neutral element for fminnum, but it isn't in the presence of NaNs :/
Good catch - something in vector legalization does that:
        t15: v4f32 = insert_vector_elt t12, ConstantFP:f32<INF>, Constant:i32<2>
      t17: v4f32 = insert_vector_elt t15, ConstantFP:f32<INF>, Constant:i32<3>
    t18: f32 = vecreduce_fmin t17



CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D87391/new/

https://reviews.llvm.org/D87391