[llvm-branch-commits] [llvm] [AMDGPU] Add wave reduce intrinsics for float types - 2 (PR #161815)

[Juan Manuel Martinez Caamaño via llvm-branch-commits]

================
@@ -5330,11 +5330,13 @@ static uint32_t getIdentityValueFor32BitWaveReduction(unsigned Opc) {
   case AMDGPU::S_MAX_U32:
     return std::numeric_limits<uint32_t>::min();
   case AMDGPU::S_MAX_I32:
+  case AMDGPU::V_SUB_F32_e64: // +0.0
----------------
jmmartinez wrote:

I haven't thought about this, but why do we take `-0.0` if the reduction is a sub and `+0.0` if it is an add ? Does it come from any specification ?

https://github.com/llvm/llvm-project/pull/161815