[llvm] [AMDGPU] Do not fold add chains wider than 32 bits into dot4 (PR #216015)

[Arseniy Obolenskiy via llvm-commits]

================
@@ -1166,6 +1166,72 @@ define i32 @dot4_and_mask10_chain_not_add(i32 %x, i32 %y) {
   ret i32 %result
 }
 
+;------------------------------------------------------------------------------
+; NEGATIVE TESTS: ACCUMULATOR WIDER THAN A DOT4 CAN REPRESENT
----------------
aobolensk wrote:

done

https://github.com/llvm/llvm-project/pull/216015