[llvm] [AMDGPU] Split wide integer dpp8 intrinsic calls (PR #113500)

[Stanislav Mekhanoshin via llvm-commits]

================
@@ -2257,6 +2260,38 @@ bool AMDGPUCodeGenPrepareImpl::visitSqrt(IntrinsicInst &Sqrt) {
   return true;
 }
 
+// Split unsupported wide integer calls.
+bool AMDGPUCodeGenPrepareImpl::visitMovDppIntrinsic(IntrinsicInst &I) {
----------------
rampitec wrote:

Didn't notice it does not happen at -O0. It looks like there is no place for it at -O0. The problem with 'other dpp intrinsics' is that it expands pseudo MI V_MOV_DPP_B64 only, past ISel, and does not handle any other types but i64. It almost calls for another pass.

https://github.com/llvm/llvm-project/pull/113500