[BACKEND] Fix workaround for unsupported cases in maybeDeduplicate (t…

…riton-lang#4048)

The logic of maybeDeduplicate doesn't seem to handle nv mma layouts. Fix
the check we have to prevent it to be used for layouts that are nv mma
or sliced of nv mma

include/triton/Conversion/TritonGPUToLLVM/ElementwiseOpToLLVMBase.h

@@ -87,10 +87,12 @@ class ElementwiseOpConversionBase : public ConvertOpToLLVMPattern<SourceOp> {
     if (!encoding)
       // encoding not available
       return resultVals;
-    if (!dyn_cast<BlockedEncodingAttr>(encoding) &&
-        !dyn_cast<SliceEncodingAttr>(encoding)) {
-      // TODO: constraining the ecndoing type here is necessary for avoiding
-      // crashes in the getElemsPerThread call below happening in the
+    Attribute baseEncoding = encoding;
+    while (auto sliced = dyn_cast<SliceEncodingAttr>(baseEncoding))
+      baseEncoding = sliced.getParent();
+    if (isa<NvidiaMmaEncodingAttr, DotOperandEncodingAttr>(baseEncoding)) {
+      // TODO: this logic seems incorrect for mma layout. Skip for now.
+      // The following test crashes and some other miscompile:
       // test_core::test_fp8_dot_acc
       return resultVals;
     }