[ElementwiseOpToLLVM] Do not call 'maybeDeduplicate' for dpas encoding (#3508)

Closes #3380 

The problem occurs when we try to lower an `arith.andi` op in a chain of
`tl.dot + tl.where + tl.dot`. The `maybeDeduplicate` function seems to
handle dot/dpas layouts incorrectly. When trying to deduplicate mask
values for each 16-col block it seems to mishandle strides & constancy
somehow leading to the mask values for the second 16-col block to be
ignored.

Signed-off-by: dchigarev <[email protected]>

Author: dchigarev

include/triton/Conversion/TritonGPUToLLVM/ElementwiseOpToLLVMBase.h

@@ -1,6 +1,7 @@
 #ifndef TRITON_CONVERSION_TRITONGPU_TO_ELEMENTWISE_OP_H
 #define TRITON_CONVERSION_TRITONGPU_TO_ELEMENTWISE_OP_H
 
+#include "intel/include/Dialect/TritonIntelGPU/IR/Attributes.h"
 #include "mlir/Conversion/LLVMCommon/Pattern.h"
 #include "mlir/Conversion/LLVMCommon/TypeConverter.h"
 #include "mlir/Support/LLVM.h"
@@ -86,7 +87,8 @@ class ElementwiseOpConversionBase : public ConvertOpToLLVMPattern<SourceOp> {
       return resultVals;
     while (auto sliced = dyn_cast<SliceEncodingAttr>(baseEncoding))
       baseEncoding = sliced.getParent();
-    if (isa<LinearEncodingAttr, DotOperandEncodingAttr>(baseEncoding)) {
+    if (isa<LinearEncodingAttr, DotOperandEncodingAttr,
+            intel::DpasEncodingAttr>(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