diff --git a/compiler/src/iree/compiler/Codegen/Common/CPU/test/llvmcpu_materialize_encoding.mlir b/compiler/src/iree/compiler/Codegen/Common/CPU/test/llvmcpu_materialize_encoding.mlir
index 821c123..466bf1f 100644
--- a/compiler/src/iree/compiler/Codegen/Common/CPU/test/llvmcpu_materialize_encoding.mlir
+++ b/compiler/src/iree/compiler/Codegen/Common/CPU/test/llvmcpu_materialize_encoding.mlir
@@ -6,10 +6,10 @@
 }{
   %c0 = arith.constant 0 : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1x1000xbf16>>
-  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1x1000xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], original_type = tensor<1x1000xbf16>, matmul_narrow_M = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], round_dims_to = array<i64: 16, 16, 16>>>>
+  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1x1000xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<1x1000xbf16>, matmul_narrow_M = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], round_dims_to = array<i64: 16, 16, 16>>>>
   %2 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1, 1000], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1x1000xbf16>> -> tensor<1x1000xbf16>
-  %3 = iree_encoding.set_encoding %2 : tensor<1x1000xbf16> -> tensor<1x1000xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], original_type = tensor<1x1000xbf16>, matmul_narrow_M = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], round_dims_to = array<i64: 16, 16, 16>>>
-  flow.dispatch.tensor.store %3, %1, offsets = [0, 0], sizes = [1, 1000], strides = [1, 1] : tensor<1x1000xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], original_type = tensor<1x1000xbf16>, matmul_narrow_M = 1  : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], round_dims_to = array<i64: 16, 16, 16>>> -> !flow.dispatch.tensor<writeonly:tensor<1x1000xbf16,  #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], original_type = tensor<1x1000xbf16>, matmul_narrow_M = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], round_dims_to = array<i64: 16, 16, 16>>>>
+  %3 = iree_encoding.set_encoding %2 : tensor<1x1000xbf16> -> tensor<1x1000xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<1x1000xbf16>, matmul_narrow_M = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], round_dims_to = array<i64: 16, 16, 16>>>
+  flow.dispatch.tensor.store %3, %1, offsets = [0, 0], sizes = [1, 1000], strides = [1, 1] : tensor<1x1000xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<1x1000xbf16>, matmul_narrow_M = 1  : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], round_dims_to = array<i64: 16, 16, 16>>> -> !flow.dispatch.tensor<writeonly:tensor<1x1000xbf16,  #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<1x1000xbf16>, matmul_narrow_M = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], round_dims_to = array<i64: 16, 16, 16>>>>
   return
 }
 // This tests that
@@ -50,7 +50,7 @@
   %8 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<7x7xf32>>
   %9 = flow.dispatch.workload.ordinal %6, 2 : index
   %10 = flow.dispatch.workload.ordinal %7, 3 : index
-  %11 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<7x7xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%9, %10}
+  %11 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<7x7xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%9, %10}
   %12 = flow.dispatch.workload.ordinal %4, 0 : index
   %13 = flow.dispatch.workload.ordinal %5, 1 : index
   %14 = flow.dispatch.tensor.load %8, offsets = [0, 0], sizes = [7, 7], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<7x7xf32>> -> tensor<7x7xf32>
@@ -60,8 +60,8 @@
   ^bb0(%arg0: index, %arg1: index):
     tensor.yield %cst : f32
   } : tensor<7x7xf32> to tensor<?x?xf32>
-  %17 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<7x7xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  flow.dispatch.tensor.store %17, %11, offsets = [0, 0], sizes = [%9, %10], strides = [1, 1] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<7x7xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<7x7xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%9, %10}
+  %17 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<7x7xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  flow.dispatch.tensor.store %17, %11, offsets = [0, 0], sizes = [%9, %10], strides = [1, 1] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<7x7xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<7x7xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%9, %10}
   return
 }
 // CHECK-LABEL:  func @set_encoding_7x7x7_matmul_LHS(
@@ -94,7 +94,7 @@
   %8 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<128x80x32xf32>>
   %9 = flow.dispatch.workload.ordinal %6, 2 : index
   %10 = flow.dispatch.workload.ordinal %7, 3 : index
-  %11 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<128x80x32xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%9, %10}
+  %11 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x80x32xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%9, %10}
   %12 = flow.dispatch.workload.ordinal %4, 0 : index
   %13 = flow.dispatch.workload.ordinal %5, 1 : index
   %14 = flow.dispatch.tensor.load %8, offsets = [0, 0, 0], sizes = [128, 80, 32], strides = [1, 1, 1] : !flow.dispatch.tensor<readonly:tensor<128x80x32xf32>> -> tensor<128x80x32xf32>
@@ -104,10 +104,10 @@
   ^bb0(%arg0: index, %arg1: index, %arg2: index):
     tensor.yield %cst : f32
   } : tensor<128x80x32xf32> to tensor<128x?x?xf32>
-  %17 = iree_encoding.set_encoding %padded : tensor<128x?x?xf32> -> tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<128x80x32xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %17 = iree_encoding.set_encoding %padded : tensor<128x?x?xf32> -> tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x80x32xf32>, user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %17, %11, offsets = [0, 0, 0], sizes = [128, %9, %10], strides = [1, 1, 1]
-    : tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<128x80x32xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-    -> !flow.dispatch.tensor<writeonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<128x80x32xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%9, %10}
+    : tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x80x32xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+    -> !flow.dispatch.tensor<writeonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x80x32xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%9, %10}
   return
 }
 // CHECK-LABEL:    func @set_encoding_128x80x32_batch_matmul_LHS(
@@ -141,7 +141,7 @@
   %10 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<128x32x320xf32>>
   %11 = flow.dispatch.workload.ordinal %8, 2 : index
   %12 = flow.dispatch.workload.ordinal %9, 3 : index
-  %13 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%5) : !flow.dispatch.tensor<writeonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<128x32x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%11, %12}
+  %13 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%5) : !flow.dispatch.tensor<writeonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x32x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%11, %12}
   %14 = flow.dispatch.workload.ordinal %6, 0 : index
   %15 = flow.dispatch.workload.ordinal %7, 1 : index
   %16 = flow.dispatch.tensor.load %10, offsets = [0, 0, 0], sizes = [128, 32, 320], strides = [1, 1, 1] : !flow.dispatch.tensor<readonly:tensor<128x32x320xf32>> -> tensor<128x32x320xf32>
@@ -151,10 +151,10 @@
   ^bb0(%arg0: index, %arg1: index, %arg2: index):
     tensor.yield %cst : f32
   } : tensor<128x32x320xf32> to tensor<128x?x?xf32>
-  %19 = iree_encoding.set_encoding %padded : tensor<128x?x?xf32> -> tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<128x32x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %19 = iree_encoding.set_encoding %padded : tensor<128x?x?xf32> -> tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x32x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %19, %13, offsets = [0, 0, 0], sizes = [128, %11, %12], strides = [1, 1, 1]
-    : tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<128x32x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-    -> !flow.dispatch.tensor<writeonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<128x32x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%11, %12}
+    : tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x32x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+    -> !flow.dispatch.tensor<writeonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x32x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%11, %12}
   return
 }
 // CHECK-LABEL:    func @set_encoding_128x32x320_batch_matmul_RHS(
@@ -183,11 +183,11 @@
   %6 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<128x80x320xf32>>
   %7 = flow.dispatch.workload.ordinal %4, 0 : index
   %8 = flow.dispatch.workload.ordinal %5, 1 : index
-  %9 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%3) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x80x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%7, %8}
+  %9 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%3) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x80x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%7, %8}
   %10 = flow.dispatch.tensor.load %9, offsets = [0, 0, 0], sizes = [128, %7, %8], strides = [1, 1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x80x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%7, %8}
-      -> tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x80x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %11 = iree_encoding.unset_encoding %10 : tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x80x320xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<128x?x?xf32>
+      : !flow.dispatch.tensor<readonly:tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x80x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>>{%7, %8}
+      -> tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x80x320xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %11 = iree_encoding.unset_encoding %10 : tensor<128x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x80x320xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<128x?x?xf32>
   %extracted_slice = tensor.extract_slice %11[0, 0, 0] [128, 80, 320] [1, 1, 1] : tensor<128x?x?xf32> to tensor<128x80x320xf32>
   flow.dispatch.tensor.store %extracted_slice, %6, offsets = [0, 0, 0], sizes = [128, 80, 320], strides = [1, 1, 1] : tensor<128x80x320xf32> -> !flow.dispatch.tensor<writeonly:tensor<128x80x320xf32>>
   return
@@ -220,14 +220,14 @@
   %cst = arith.constant 0.0 : f32
   %d0 = tensor.dim %arg0, %c0 : tensor<?x?xf32>
   %d1 = tensor.dim %arg1, %c1 : tensor<?x?xf32>
-  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %2 = tensor.empty(%d0, %d1) : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %3 = linalg.fill ins(%cst : f32) outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %4 = linalg.matmul ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %5 = iree_encoding.unset_encoding %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
+  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %2 = tensor.empty(%d0, %d1) : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %3 = linalg.fill ins(%cst : f32) outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %4 = linalg.matmul ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %5 = iree_encoding.unset_encoding %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
   return %5 : tensor<?x?xf32>
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -269,19 +269,19 @@
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %cst : f32
   } : tensor<16x16xf32> to tensor<?x?xf32>
-  %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<16x16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
+  %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16x16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
   %padded_0 = tensor.pad %1 low[0, 0] high[%c0, %c0] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %cst : f32
   } : tensor<16x1xf32> to tensor<?x?xf32>
-  %4 = iree_encoding.set_encoding %padded_0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
+  %4 = iree_encoding.set_encoding %padded_0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
   %padded_1 = tensor.pad %2 low[0, 0] high[%c0, %c0] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %cst : f32
   } : tensor<16x1xf32> to tensor<?x?xf32>
-  %5 = iree_encoding.set_encoding %padded_1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
-  %6 = linalg.matmul ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<16x16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>) outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
-  %7 = iree_encoding.unset_encoding %6 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>> -> tensor<?x?xf32>
+  %5 = iree_encoding.set_encoding %padded_1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
+  %6 = linalg.matmul ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16x16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>) outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
+  %7 = iree_encoding.unset_encoding %6 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16x1xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>> -> tensor<?x?xf32>
   %extracted_slice = tensor.extract_slice %7[0, 0] [16, 1] [1, 1] : tensor<?x?xf32> to tensor<16x1xf32>
   %8 = hal.tensor.export %extracted_slice "output0" : tensor<16x1xf32> -> !hal.buffer_view
   func.return %8 : !hal.buffer_view
@@ -304,28 +304,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -371,19 +371,19 @@
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %cst : f32
   } : tensor<16x16xf32> to tensor<?x?xf32>
-  %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<16x16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>
+  %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16x16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>
   %padded_0 = tensor.pad %1 low[0] high[%c0] {
   ^bb0(%arg3: index):
     tensor.yield %cst : f32
   } : tensor<16xf32> to tensor<?xf32>
-  %4 = iree_encoding.set_encoding %padded_0 : tensor<?xf32> -> tensor<?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>
+  %4 = iree_encoding.set_encoding %padded_0 : tensor<?xf32> -> tensor<?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>
   %padded_1 = tensor.pad %2 low[0] high[%c0] {
   ^bb0(%arg3: index):
     tensor.yield %cst : f32
   } : tensor<16xf32> to tensor<?xf32>
-  %5 = iree_encoding.set_encoding %padded_1 : tensor<?xf32> -> tensor<?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>
-  %6 = linalg.matvec ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<16x16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>, tensor<?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>) outs(%5 : tensor<?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>) -> tensor<?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>
-  %7 = iree_encoding.unset_encoding %6 : tensor<?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>> -> tensor<?xf32>
+  %5 = iree_encoding.set_encoding %padded_1 : tensor<?xf32> -> tensor<?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>
+  %6 = linalg.matvec ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16x16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>, tensor<?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>) outs(%5 : tensor<?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>) -> tensor<?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>>
+  %7 = iree_encoding.unset_encoding %6 : tensor<?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<16xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0)>]>> -> tensor<?xf32>
   %extracted_slice = tensor.extract_slice %7[0] [16] [1] : tensor<?xf32> to tensor<16xf32>
   %8 = hal.tensor.export %extracted_slice "output0" : tensor<16xf32> -> !hal.buffer_view
   func.return %8 : !hal.buffer_view
@@ -403,28 +403,28 @@
 } {
   %c0 = arith.constant 0 : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<16x16xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
+      : !flow.dispatch.tensor<readonly:tensor<16x16xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<16x1xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
+      : !flow.dispatch.tensor<readonly:tensor<16x1xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
+      : !flow.dispatch.tensor<readwrite:tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [16, 16], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<16x16xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
-      -> tensor<16x16xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<16x16xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
+      -> tensor<16x16xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [16, 1], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<16x1xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
-      -> tensor<16x1xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<16x1xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
+      -> tensor<16x1xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [16, 1], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
-      -> tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
+      -> tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<16x16xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<16x1xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<16x16xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<16x1xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [16, 1], strides = [1, 1]
-      : tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
+      : tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<16x1xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>>
   return
 }
 // CHECK-LABEL: func @matvec_lowering_f32f32f32_aarch64()
@@ -460,28 +460,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -523,28 +523,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -587,28 +587,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -650,28 +650,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 16)>
@@ -713,28 +713,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 16)>
@@ -776,28 +776,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 16)>
@@ -839,28 +839,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 16)>
@@ -902,28 +902,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 16)>
@@ -965,28 +965,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 16)>
@@ -1030,28 +1030,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 16)>
@@ -1095,37 +1095,37 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %lhs_f32 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %rhs = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %dest = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
 
-  %empty = tensor.empty(%M, %K) : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+  %empty = tensor.empty(%M, %K) : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %lhs_f16 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]}
-     ins(%lhs_f32 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
-     outs(%empty : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>) {
+     ins(%lhs_f32 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
+     outs(%empty : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: f32, %out: f16):
     %17 = arith.truncf %in : f32 to f16
     linalg.yield %17 : f16
-  } -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+  } -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%lhs_f16, %rhs : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%dest : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%lhs_f16, %rhs : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%dest : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP_CEILDIV_8:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -1160,37 +1160,37 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %lhs_f32 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %rhs = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %dest = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
 
-  %empty = tensor.empty(%M, %K) : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+  %empty = tensor.empty(%M, %K) : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %lhs_f16 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]}
-     ins(%lhs_f32 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
-     outs(%empty : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>) {
+     ins(%lhs_f32 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
+     outs(%empty : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: f32, %out: f16):
     %17 = arith.truncf %in : f32 to f16
     linalg.yield %17 : f16
-  } -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+  } -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%lhs_f16, %rhs : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%dest : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%lhs_f16, %rhs : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%dest : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f16, f16], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 
@@ -1226,28 +1226,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 // CHECK-LABEL: func @matmul_lowering_i8i8i32_aarch64()
@@ -1286,28 +1286,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -1351,28 +1351,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -1415,28 +1415,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 4)>
@@ -1481,28 +1481,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -1545,28 +1545,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 4)>
@@ -1606,15 +1606,15 @@
 func.func @matmul_lowering_f32f32f32_aarch64_sve(%lhs : tensor<?x?xf32>, %rhs: tensor<?x?xf32>, %acc: tensor<?x?xf32>) -> tensor<?x?xf32> attributes {
   hal.executable.target = #hal.executable.target<"xyz", "xyz", {cpu_features = "+sve", target_triple="aarch64-xyz-xyz"}>
 } {
-  %0 = iree_encoding.set_encoding %lhs : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %1 = iree_encoding.set_encoding %rhs : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %2 = iree_encoding.set_encoding %acc : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %0 = iree_encoding.set_encoding %lhs : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %1 = iree_encoding.set_encoding %rhs : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %2 = iree_encoding.set_encoding %acc : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %3 = linalg.matmul
-      ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
+      ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
   return %4 : tensor<?x?xf32>
 }
 
@@ -1631,15 +1631,15 @@
 func.func @matmul_lowering_f32f32f32_riscv(%lhs : tensor<?x?xf32>, %rhs: tensor<?x?xf32>, %acc: tensor<?x?xf32>) -> tensor<?x?xf32> attributes {
   hal.executable.target = #hal.executable.target<"xyz", "xyz", {target_triple="riscv32-xyz-xyz"}>
 } {
-  %0 = iree_encoding.set_encoding %lhs : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %1 = iree_encoding.set_encoding %rhs : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %2 = iree_encoding.set_encoding %acc : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %0 = iree_encoding.set_encoding %lhs : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %1 = iree_encoding.set_encoding %rhs : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %2 = iree_encoding.set_encoding %acc : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %3 = linalg.matmul
-      ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
+      ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
   return %4 : tensor<?x?xf32>
 }
 // RISC-V targets does not implement data-tiling yet.
@@ -1660,28 +1660,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -1725,28 +1725,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -1790,28 +1790,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 16)>
@@ -1855,28 +1855,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 16)>
@@ -1926,28 +1926,28 @@
   ^bb0(%arg2: index, %arg3: index, %arg4: index):
     tensor.yield %c0_i8 : i8
   } : tensor<32x1x128xi8> to tensor<?x?x?xi8>
-  %4 = iree_encoding.set_encoding %padded : tensor<?x?x?xi8> -> tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %5 = tensor.empty(%c32, %c1, %c128) : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %6 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>], iterator_types = ["parallel", "parallel", "parallel"]} ins(%4 : tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%5 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
+  %4 = iree_encoding.set_encoding %padded : tensor<?x?x?xi8> -> tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %5 = tensor.empty(%c32, %c1, %c128) : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %6 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>], iterator_types = ["parallel", "parallel", "parallel"]} ins(%4 : tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%5 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i8, %out: i32):
     %17 = arith.extsi %in : i8 to i32
     linalg.yield %17 : i32
-  } -> tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  } -> tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
   %padded_0 = tensor.pad %arg1 low[0, 0, 0] high[%c0, %c0, %c0] {
   ^bb0(%arg2: index, %arg3: index, %arg4: index):
     tensor.yield %c0_i8 : i8
   } : tensor<32x128x11008xi8> to tensor<?x?x?xi8>
-  %7 = iree_encoding.set_encoding %padded_0 : tensor<?x?x?xi8> -> tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %8 = tensor.empty(%c32, %c128, %c11008) : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %9 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>], iterator_types = ["parallel", "parallel", "parallel"]} ins(%7 : tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%8 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
+  %7 = iree_encoding.set_encoding %padded_0 : tensor<?x?x?xi8> -> tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %8 = tensor.empty(%c32, %c128, %c11008) : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %9 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>], iterator_types = ["parallel", "parallel", "parallel"]} ins(%7 : tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%8 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i8, %out: i32):
     %17 = arith.extsi %in : i8 to i32
     linalg.yield %17 : i32
-  } -> tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %10 = tensor.empty(%c32, %c1, %c11008) : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %11 = linalg.fill ins(%c0_i32 : i32) outs(%10 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %12 = linalg.batch_matmul ins(%6, %9 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%11 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %13 = iree_encoding.unset_encoding %12 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?x?xi32>
+  } -> tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %10 = tensor.empty(%c32, %c1, %c11008) : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %11 = linalg.fill ins(%c0_i32 : i32) outs(%10 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %12 = linalg.batch_matmul ins(%6, %9 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%11 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %13 = iree_encoding.unset_encoding %12 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x1x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?x?xi32>
   %extracted_slice = tensor.extract_slice %13[0, 0, 0] [32, 1, 11008] [1, 1, 1] : tensor<?x?x?xi32> to tensor<32x1x11008xi32>
   return %extracted_slice : tensor<32x1x11008xi32>
 }
@@ -1990,28 +1990,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 1, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 1, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi16, #iree_encoding.encoding<operand_index = 1, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi16, #iree_encoding.encoding<operand_index = 1, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi16, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, i16, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //   CHECK-DAG: #[[$MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -2055,35 +2055,35 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %lhs_binding = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %rhs_binding = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %out_binding = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %lhs = flow.dispatch.tensor.load %lhs_binding, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %rhs_i4 = flow.dispatch.tensor.load %rhs_binding, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
-  %empty = tensor.empty(%K, %N) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+  %empty = tensor.empty(%K, %N) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %rhs_i32 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]}
-     ins(%rhs_i4 : tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>) outs(%empty : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>) {
+     ins(%rhs_i4 : tensor<?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>) outs(%empty : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i4, %out: i32):
     %17 = arith.extui %in : i4 to i32
     linalg.yield %17 : i32
-  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %out = flow.dispatch.tensor.load %out_binding, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %result = linalg.matmul
-      ins(%lhs, %rhs_i32 : tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%out : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%lhs, %rhs_i32 : tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%out : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %result, %out_binding, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 
@@ -2125,28 +2125,28 @@
   ^bb0(%arg2: index):
     tensor.yield %c0_i8 : i8
   } : tensor<128xi8> to tensor<?xi8>
-  %4 = iree_encoding.set_encoding %padded : tensor<?xi8> -> tensor<?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %5 = tensor.empty(%c128) : tensor<?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %6 = linalg.generic {indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> (d0)>], iterator_types = ["parallel"]} ins(%4 : tensor<?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%5 : tensor<?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
+  %4 = iree_encoding.set_encoding %padded : tensor<?xi8> -> tensor<?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %5 = tensor.empty(%c128) : tensor<?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %6 = linalg.generic {indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> (d0)>], iterator_types = ["parallel"]} ins(%4 : tensor<?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%5 : tensor<?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i8, %out: i32):
     %17 = arith.extsi %in : i8 to i32
     linalg.yield %17 : i32
-  } -> tensor<?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  } -> tensor<?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
   %padded_0 = tensor.pad %arg1 low[0, 0] high[%c0, %c0] {
   ^bb0(%arg2: index, %arg3: index):
     tensor.yield %c0_i8 : i8
   } : tensor<128x11008xi8> to tensor<?x?xi8>
-  %7 = iree_encoding.set_encoding %padded_0 : tensor<?x?xi8> -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %8 = tensor.empty(%c128, %c11008) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %9 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%7 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%8 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
+  %7 = iree_encoding.set_encoding %padded_0 : tensor<?x?xi8> -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %8 = tensor.empty(%c128, %c11008) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %9 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%7 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%8 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i8, %out: i32):
     %17 = arith.extsi %in : i8 to i32
     linalg.yield %17 : i32
-  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %10 = tensor.empty(%c11008) : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %11 = linalg.fill ins(%c0_i32 : i32) outs(%10 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %12 = linalg.vecmat ins(%6, %9 : tensor<?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%11 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %13 = iree_encoding.unset_encoding %12 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?xi32>
+  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %10 = tensor.empty(%c11008) : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %11 = linalg.fill ins(%c0_i32 : i32) outs(%10 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %12 = linalg.vecmat ins(%6, %9 : tensor<?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%11 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %13 = iree_encoding.unset_encoding %12 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?xi32>
   %extracted_slice = tensor.extract_slice %13[0] [11008] [1] : tensor<?xi32> to tensor<11008xi32>
   return %extracted_slice : tensor<11008xi32>
 }
@@ -2198,28 +2198,28 @@
   ^bb0(%arg2: index, %arg3: index):
     tensor.yield %c0_i8 : i8
   } : tensor<11008x128xi8> to tensor<?x?xi8>
-  %4 = iree_encoding.set_encoding %padded : tensor<?x?xi8> -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %5 = tensor.empty(%c11008, %c128) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %6 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
+  %4 = iree_encoding.set_encoding %padded : tensor<?x?xi8> -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %5 = tensor.empty(%c11008, %c128) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %6 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i8, %out: i32):
     %17 = arith.extsi %in : i8 to i32
     linalg.yield %17 : i32
-  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
   %padded_0 = tensor.pad %arg1 low[0] high[%c0] {
   ^bb0(%arg2: index):
     tensor.yield %c0_i8 : i8
   } : tensor<128xi8> to tensor<?xi8>
-  %7 = iree_encoding.set_encoding %padded_0 : tensor<?xi8> -> tensor<?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %8 = tensor.empty(%c128) : tensor<?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %9 = linalg.generic {indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> (d0)>], iterator_types = ["parallel"]} ins(%7 : tensor<?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%8 : tensor<?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
+  %7 = iree_encoding.set_encoding %padded_0 : tensor<?xi8> -> tensor<?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %8 = tensor.empty(%c128) : tensor<?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %9 = linalg.generic {indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> (d0)>], iterator_types = ["parallel"]} ins(%7 : tensor<?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%8 : tensor<?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i8, %out: i32):
     %17 = arith.extsi %in : i8 to i32
     linalg.yield %17 : i32
-  } -> tensor<?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %10 = tensor.empty(%c11008) : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %11 = linalg.fill ins(%c0_i32 : i32) outs(%10 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %12 = linalg.matvec ins(%6, %9 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%11 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %13 = iree_encoding.unset_encoding %12 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?xi32>
+  } -> tensor<?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %10 = tensor.empty(%c11008) : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %11 = linalg.fill ins(%c0_i32 : i32) outs(%10 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %12 = linalg.matvec ins(%6, %9 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%11 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %13 = iree_encoding.unset_encoding %12 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_N = 1 : index, original_type = tensor<11008xi32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?xi32>
   %extracted_slice = tensor.extract_slice %13[0] [11008] [1] : tensor<?xi32> to tensor<11008xi32>
   return %extracted_slice : tensor<11008xi32>
 }
@@ -2271,28 +2271,28 @@
   ^bb0(%arg2: index, %arg3: index):
     tensor.yield %c0_i8 : i8
   } : tensor<15x128xi8> to tensor<?x?xi8>
-  %4 = iree_encoding.set_encoding %padded : tensor<?x?xi8> -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %5 = tensor.empty(%c15, %c128) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %6 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
+  %4 = iree_encoding.set_encoding %padded : tensor<?x?xi8> -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %5 = tensor.empty(%c15, %c128) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %6 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i8, %out: i32):
     %17 = arith.extsi %in : i8 to i32
     linalg.yield %17 : i32
-  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
   %padded_0 = tensor.pad %arg1 low[0] high[%c0] {
   ^bb0(%arg2: index):
     tensor.yield %c0_i8 : i8
   } : tensor<128xi8> to tensor<?xi8>
-  %7 = iree_encoding.set_encoding %padded_0 : tensor<?xi8> -> tensor<?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %8 = tensor.empty(%c128) : tensor<?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %9 = linalg.generic {indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> (d0)>], iterator_types = ["parallel"]} ins(%7 : tensor<?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%8 : tensor<?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
+  %7 = iree_encoding.set_encoding %padded_0 : tensor<?xi8> -> tensor<?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %8 = tensor.empty(%c128) : tensor<?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %9 = linalg.generic {indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> (d0)>], iterator_types = ["parallel"]} ins(%7 : tensor<?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%8 : tensor<?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i8, %out: i32):
     %17 = arith.extsi %in : i8 to i32
     linalg.yield %17 : i32
-  } -> tensor<?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %10 = tensor.empty(%c15) : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %11 = linalg.fill ins(%c0_i32 : i32) outs(%10 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %12 = linalg.matvec ins(%6, %9 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%11 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %13 = iree_encoding.unset_encoding %12 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?xi32>
+  } -> tensor<?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %10 = tensor.empty(%c15) : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %11 = linalg.fill ins(%c0_i32 : i32) outs(%10 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %12 = linalg.matvec ins(%6, %9 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%11 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %13 = iree_encoding.unset_encoding %12 : tensor<?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 15 : index, matmul_narrow_N = 1 : index, original_type = tensor<15xi32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?xi32>
   %extracted_slice = tensor.extract_slice %13[0] [15] [1] : tensor<?xi32> to tensor<15xi32>
   return %extracted_slice : tensor<15xi32>
 }
@@ -2346,28 +2346,28 @@
   ^bb0(%arg2: index, %arg3: index):
     tensor.yield %c0_i8 : i8
   } : tensor<32x128xi8> to tensor<?x?xi8>
-  %4 = iree_encoding.set_encoding %padded : tensor<?x?xi8> -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %5 = tensor.empty(%c32, %c128) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %6 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
+  %4 = iree_encoding.set_encoding %padded : tensor<?x?xi8> -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %5 = tensor.empty(%c32, %c128) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %6 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i8, %out: i32):
     %17 = arith.extsi %in : i8 to i32
     linalg.yield %17 : i32
-  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>
   %padded_0 = tensor.pad %arg1 low[0, 0, 0] high[%c0, %c0, %c0] {
   ^bb0(%arg2: index, %arg3: index, %arg4: index):
     tensor.yield %c0_i8 : i8
   } : tensor<32x128x11008xi8> to tensor<?x?x?xi8>
-  %7 = iree_encoding.set_encoding %padded_0 : tensor<?x?x?xi8> -> tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %8 = tensor.empty(%c32, %c128, %c11008) : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %9 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>], iterator_types = ["parallel", "parallel", "parallel"]} ins(%7 : tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%8 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
+  %7 = iree_encoding.set_encoding %padded_0 : tensor<?x?x?xi8> -> tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %8 = tensor.empty(%c32, %c128, %c11008) : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %9 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>], iterator_types = ["parallel", "parallel", "parallel"]} ins(%7 : tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%8 : tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i8, %out: i32):
     %17 = arith.extsi %in : i8 to i32
     linalg.yield %17 : i32
-  } -> tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
-  %10 = tensor.empty(%c32, %c11008) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %11 = linalg.fill ins(%c0_i32 : i32) outs(%10 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %12 = linalg.batch_vecmat ins(%6, %9 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%11 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %13 = iree_encoding.unset_encoding %12 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xi32>
+  } -> tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>
+  %10 = tensor.empty(%c32, %c11008) : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %11 = linalg.fill ins(%c0_i32 : i32) outs(%10 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %12 = linalg.batch_vecmat ins(%6, %9 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128xi8>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x128x11008xi8>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%11 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %13 = iree_encoding.unset_encoding %12 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], matmul_narrow_M = 1 : index, original_type = tensor<32x11008xi32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xi32>
   %extracted_slice = tensor.extract_slice %13[0, 0] [32, 11008] [1, 1] : tensor<?x?xi32> to tensor<32x11008xi32>
   return %extracted_slice : tensor<32x11008xi32>
 }
@@ -2416,19 +2416,19 @@
   ^bb0(%arg3: index, %arg4: index, %arg5: index):
     tensor.yield %c0_i8 : i8
   } : tensor<32x11008x128xi8> to tensor<?x?x?xi8>
-  %3 = iree_encoding.set_encoding %padded : tensor<?x?x?xi8> -> tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], original_type = tensor<32x11008x128xi8>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
+  %3 = iree_encoding.set_encoding %padded : tensor<?x?x?xi8> -> tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<32x11008x128xi8>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
   %padded_0 = tensor.pad %1 low[0, 0] high[%c0, %c0] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %c0_i8 : i8
   } : tensor<32x128xi8> to tensor<?x?xi8>
-  %4 = iree_encoding.set_encoding %padded_0 : tensor<?x?xi8> -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], original_type = tensor<32x128xi8>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
+  %4 = iree_encoding.set_encoding %padded_0 : tensor<?x?xi8> -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<32x128xi8>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
   %padded_1 = tensor.pad %2 low[0, 0] high[%c0, %c0] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %c0_i32 : i32
   } : tensor<32x11008xi32> to tensor<?x?xi32>
-  %5 = iree_encoding.set_encoding %padded_1 : tensor<?x?xi32> -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], original_type = tensor<32x11008xi32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
-  %6 = linalg.batch_matvec ins(%3, %4 : tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], original_type = tensor<32x11008x128xi8>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>, tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], original_type = tensor<32x128xi8>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>) outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], original_type = tensor<32x11008xi32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>) -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], original_type = tensor<32x11008xi32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
-  %7 = iree_encoding.unset_encoding %6 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], original_type = tensor<32x11008xi32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>> -> tensor<?x?xi32>
+  %5 = iree_encoding.set_encoding %padded_1 : tensor<?x?xi32> -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<32x11008xi32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
+  %6 = linalg.batch_matvec ins(%3, %4 : tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<32x11008x128xi8>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>, tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<32x128xi8>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>) outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<32x11008xi32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>) -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<32x11008xi32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
+  %7 = iree_encoding.unset_encoding %6 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<32x11008xi32>, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>> -> tensor<?x?xi32>
   %extracted_slice = tensor.extract_slice %7[0, 0] [32, 11008] [1, 1] : tensor<?x?xi32> to tensor<32x11008xi32>
   %8 = hal.tensor.export %extracted_slice "output0" : tensor<32x11008xi32> -> !hal.buffer_view
   func.return %8 : !hal.buffer_view
@@ -2449,32 +2449,32 @@
   %c128 = arith.constant 128 : index
   %cst = arith.constant 0.000000e+00 : f32
   %c512 = arith.constant 512 : index
-  %3:2 = iree_encoding.upper_bound_tile_size tensor<256x128xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %3:2 = iree_encoding.upper_bound_tile_size tensor<256x128xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %4 = affine.apply #map3()[%3#0, %c256]
   %5 = affine.apply #map3()[%3#1, %c128]
   %padded = tensor.pad %arg0 low[0, 0] high[%4, %5] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %cst : f32
   } : tensor<256x128xf32> to tensor<?x?xf32>
-  %6 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<256x128xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %7:2 = iree_encoding.upper_bound_tile_size tensor<256x512xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %6 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<256x128xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %7:2 = iree_encoding.upper_bound_tile_size tensor<256x512xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %8 = affine.apply #map3()[%7#0, %c256]
   %9 = affine.apply #map3()[%7#1, %c512]
   %padded_0 = tensor.pad %arg1 low[0, 0] high[%8, %9] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %cst : f32
   } : tensor<256x512xf32> to tensor<?x?xf32>
-  %10 = iree_encoding.set_encoding %padded_0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<256x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %11:2 = iree_encoding.upper_bound_tile_size tensor<128x512xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %10 = iree_encoding.set_encoding %padded_0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<256x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %11:2 = iree_encoding.upper_bound_tile_size tensor<128x512xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %12 = affine.apply #map3()[%11#0, %c128]
   %13 = affine.apply #map3()[%11#1, %c512]
   %padded_1 = tensor.pad %arg2 low[0, 0] high[%12, %13] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %cst : f32
   } : tensor<128x512xf32> to tensor<?x?xf32>
-  %14 = iree_encoding.set_encoding %padded_1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %15 = linalg.matmul_transpose_a ins(%6, %10 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<256x128xf32>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<256x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%14 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %16 = iree_encoding.unset_encoding %15 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
+  %14 = iree_encoding.set_encoding %padded_1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %15 = linalg.matmul_transpose_a ins(%6, %10 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<256x128xf32>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<256x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%14 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %16 = iree_encoding.unset_encoding %15 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
   %extracted_slice = tensor.extract_slice %16[0, 0] [128, 512] [1, 1] : tensor<?x?xf32> to tensor<128x512xf32>
   return %extracted_slice : tensor<128x512xf32>
 }
@@ -2507,32 +2507,32 @@
   %c256 = arith.constant 256 : index
   %cst = arith.constant 0.000000e+00 : f32
   %c512 = arith.constant 512 : index
-  %3:2 = iree_encoding.upper_bound_tile_size tensor<128x256xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %3:2 = iree_encoding.upper_bound_tile_size tensor<128x256xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %4 = affine.apply #map3()[%3#0, %c128]
   %5 = affine.apply #map3()[%3#1, %c256]
   %padded = tensor.pad %arg0 low[0, 0] high[%4, %5] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %cst : f32
   } : tensor<128x256xf32> to tensor<?x?xf32>
-  %6 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<128x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %7:2 = iree_encoding.upper_bound_tile_size tensor<512x256xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %6 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %7:2 = iree_encoding.upper_bound_tile_size tensor<512x256xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %8 = affine.apply #map3()[%7#0, %c512]
   %9 = affine.apply #map3()[%7#1, %c256]
   %padded_0 = tensor.pad %arg1 low[0, 0] high[%8, %9] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %cst : f32
   } : tensor<512x256xf32> to tensor<?x?xf32>
-  %10 = iree_encoding.set_encoding %padded_0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<512x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %11:2 = iree_encoding.upper_bound_tile_size tensor<128x512xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %10 = iree_encoding.set_encoding %padded_0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<512x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %11:2 = iree_encoding.upper_bound_tile_size tensor<128x512xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %12 = affine.apply #map3()[%11#0, %c128]
   %13 = affine.apply #map3()[%11#1, %c512]
   %padded_1 = tensor.pad %arg2 low[0, 0] high[%12, %13] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %cst : f32
   } : tensor<128x512xf32> to tensor<?x?xf32>
-  %14 = iree_encoding.set_encoding %padded_1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %15 = linalg.matmul_transpose_b ins(%6, %10 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<128x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<512x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%14 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %16 = iree_encoding.unset_encoding %15 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
+  %14 = iree_encoding.set_encoding %padded_1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %15 = linalg.matmul_transpose_b ins(%6, %10 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<512x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%14 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %16 = iree_encoding.unset_encoding %15 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
   %extracted_slice = tensor.extract_slice %16[0, 0] [128, 512] [1, 1] : tensor<?x?xf32> to tensor<128x512xf32>
   return %extracted_slice : tensor<128x512xf32>
 }
@@ -2565,7 +2565,7 @@
   %c128 = arith.constant 128 : index
   %cst = arith.constant 0.000000e+00 : f32
   %c512 = arith.constant 512 : index
-  %3:3 = iree_encoding.upper_bound_tile_size tensor<2x256x128xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
+  %3:3 = iree_encoding.upper_bound_tile_size tensor<2x256x128xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
   %4 = affine.apply #map3()[%3#0, %c2]
   %5 = affine.apply #map3()[%3#1, %c256]
   %6 = affine.apply #map3()[%3#2, %c128]
@@ -2573,8 +2573,8 @@
   ^bb0(%arg3: index, %arg4: index, %arg5: index):
     tensor.yield %cst : f32
   } : tensor<2x256x128xf32> to tensor<?x?x?xf32>
-  %7 = iree_encoding.set_encoding %padded : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<2x256x128xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %8:3 = iree_encoding.upper_bound_tile_size tensor<2x256x512xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
+  %7 = iree_encoding.set_encoding %padded : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x256x128xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %8:3 = iree_encoding.upper_bound_tile_size tensor<2x256x512xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
   %9 = affine.apply #map3()[%8#0, %c2]
   %10 = affine.apply #map3()[%8#1, %c256]
   %11 = affine.apply #map3()[%8#2, %c512]
@@ -2582,8 +2582,8 @@
   ^bb0(%arg3: index, %arg4: index, %arg5: index):
     tensor.yield %cst : f32
   } : tensor<2x256x512xf32> to tensor<?x?x?xf32>
-  %12 = iree_encoding.set_encoding %padded_0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<2x256x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %13:3 = iree_encoding.upper_bound_tile_size tensor<2x128x512xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
+  %12 = iree_encoding.set_encoding %padded_0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x256x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %13:3 = iree_encoding.upper_bound_tile_size tensor<2x128x512xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
   %14 = affine.apply #map3()[%13#0, %c2]
   %15 = affine.apply #map3()[%13#1, %c128]
   %16 = affine.apply #map3()[%13#2, %c512]
@@ -2591,9 +2591,9 @@
   ^bb0(%arg3: index, %arg4: index, %arg5: index):
     tensor.yield %cst : f32
   } : tensor<2x128x512xf32> to tensor<?x?x?xf32>
-  %17 = iree_encoding.set_encoding %padded_1 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %18 = linalg.batch_matmul_transpose_a ins(%7, %12 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<2x256x128xf32>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<2x256x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%17 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %19 = iree_encoding.unset_encoding %18 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?x?xf32>
+  %17 = iree_encoding.set_encoding %padded_1 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %18 = linalg.batch_matmul_transpose_a ins(%7, %12 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x256x128xf32>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x256x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%17 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %19 = iree_encoding.unset_encoding %18 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?x?xf32>
   %extracted_slice = tensor.extract_slice %19[0, 0, 0] [2, 128, 512] [1, 1, 1] : tensor<?x?x?xf32> to tensor<2x128x512xf32>
   return %extracted_slice : tensor<2x128x512xf32>
 }
@@ -2626,7 +2626,7 @@
   %c256 = arith.constant 256 : index
   %cst = arith.constant 0.000000e+00 : f32
   %c512 = arith.constant 512 : index
-  %3:3 = iree_encoding.upper_bound_tile_size tensor<2x128x256xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
+  %3:3 = iree_encoding.upper_bound_tile_size tensor<2x128x256xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
   %4 = affine.apply #map3()[%3#0, %c2]
   %5 = affine.apply #map3()[%3#1, %c128]
   %6 = affine.apply #map3()[%3#2, %c256]
@@ -2634,8 +2634,8 @@
   ^bb0(%arg3: index, %arg4: index, %arg5: index):
     tensor.yield %cst : f32
   } : tensor<2x128x256xf32> to tensor<?x?x?xf32>
-  %7 = iree_encoding.set_encoding %padded : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<2x128x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %8:3 = iree_encoding.upper_bound_tile_size tensor<2x512x256xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
+  %7 = iree_encoding.set_encoding %padded : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x128x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %8:3 = iree_encoding.upper_bound_tile_size tensor<2x512x256xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
   %9 = affine.apply #map3()[%8#0, %c2]
   %10 = affine.apply #map3()[%8#1, %c512]
   %11 = affine.apply #map3()[%8#2, %c256]
@@ -2643,8 +2643,8 @@
   ^bb0(%arg3: index, %arg4: index, %arg5: index):
     tensor.yield %cst : f32
   } : tensor<2x512x256xf32> to tensor<?x?x?xf32>
-  %12 = iree_encoding.set_encoding %padded_0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<2x512x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %13:3 = iree_encoding.upper_bound_tile_size tensor<2x128x512xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
+  %12 = iree_encoding.set_encoding %padded_0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x512x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %13:3 = iree_encoding.upper_bound_tile_size tensor<2x128x512xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
   %14 = affine.apply #map3()[%13#0, %c2]
   %15 = affine.apply #map3()[%13#1, %c128]
   %16 = affine.apply #map3()[%13#2, %c512]
@@ -2652,9 +2652,9 @@
   ^bb0(%arg3: index, %arg4: index, %arg5: index):
     tensor.yield %cst : f32
   } : tensor<2x128x512xf32> to tensor<?x?x?xf32>
-  %17 = iree_encoding.set_encoding %padded_1 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %18 = linalg.batch_matmul_transpose_b ins(%7, %12 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<2x128x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<2x512x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%17 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
-  %19 = iree_encoding.unset_encoding %18 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?x?xf32>
+  %17 = iree_encoding.set_encoding %padded_1 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %18 = linalg.batch_matmul_transpose_b ins(%7, %12 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x128x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x512x256xf32>, user_indexing_maps = [#map, #map1, #map2]>>) outs(%17 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>>
+  %19 = iree_encoding.unset_encoding %18 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x128x512xf32>, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?x?xf32>
   %extracted_slice = tensor.extract_slice %19[0, 0, 0] [2, 128, 512] [1, 1, 1] : tensor<?x?x?xf32> to tensor<2x128x512xf32>
   return %extracted_slice : tensor<2x128x512xf32>
 }
@@ -2688,15 +2688,15 @@
   %c0_i16 = arith.constant 0 : i16
   %c128 = arith.constant 128 : index
   %c32 = arith.constant 32 : index
-  %0:2 = iree_encoding.upper_bound_tile_size tensor<32x128xi16, #iree_encoding.encoding<operand_index = 0, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %0:2 = iree_encoding.upper_bound_tile_size tensor<32x128xi16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %1 = affine.apply #map3()[%0#0, %c32]
   %2 = affine.apply #map3()[%0#1, %c128]
   %padded = tensor.pad %arg0 low[0, 0] high[%1, %2] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %c0_i16 : i16
   } : tensor<32x128xi16> to tensor<?x?xi16>
-  %3 = iree_encoding.set_encoding %padded : tensor<?x?xi16> -> tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, element_types = [i16, ui4, i32], original_type = tensor<32x128xi16>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
-  %4:3 = iree_encoding.upper_bound_tile_size tensor<4096x32x128xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
+  %3 = iree_encoding.set_encoding %padded : tensor<?x?xi16> -> tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i16, ui4, i32], original_type = tensor<32x128xi16>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
+  %4:3 = iree_encoding.upper_bound_tile_size tensor<4096x32x128xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>> -> index, index, index
   %5 = affine.apply #map3()[%4#0, %c4096]
   %6 = affine.apply #map3()[%4#1, %c32]
   %7 = affine.apply #map3()[%4#2, %c128]
@@ -2704,24 +2704,24 @@
   ^bb0(%arg3: index, %arg4: index, %arg5: index):
     tensor.yield %c0_i4 : i4
   } : tensor<4096x32x128xi4> to tensor<?x?x?xi4>
-  %8 = iree_encoding.set_encoding %padded_0 : tensor<?x?x?xi4> -> tensor<?x?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i16, ui4, i32], original_type = tensor<4096x32x128xi4>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
-  %9:2 = iree_encoding.upper_bound_tile_size tensor<4096x32xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %8 = iree_encoding.set_encoding %padded_0 : tensor<?x?x?xi4> -> tensor<?x?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, ui4, i32], original_type = tensor<4096x32x128xi4>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
+  %9:2 = iree_encoding.upper_bound_tile_size tensor<4096x32xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %10 = affine.apply #map3()[%9#0, %c4096]
   %11 = affine.apply #map3()[%9#1, %c32]
   %padded_1 = tensor.pad %arg2 low[0, 0] high[%10, %11] {
   ^bb0(%arg3: index, %arg4: index):
     tensor.yield %c0_i32 : i32
   } : tensor<4096x32xi32> to tensor<?x?xi32>
-  %12 = iree_encoding.set_encoding %padded_1 : tensor<?x?xi32> -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], original_type = tensor<4096x32xi32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
-  %13 = linalg.generic {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} ins(%3, %8 : tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, element_types = [i16, ui4, i32], original_type = tensor<32x128xi16>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xi4, #iree_encoding.encoding<operand_index = 1, element_types = [i16, ui4, i32], original_type = tensor<4096x32x128xi4>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>) outs(%12 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], original_type = tensor<4096x32xi32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>) {
+  %12 = iree_encoding.set_encoding %padded_1 : tensor<?x?xi32> -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], original_type = tensor<4096x32xi32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
+  %13 = linalg.generic {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} ins(%3, %8 : tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i16, ui4, i32], original_type = tensor<32x128xi16>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xi4, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i16, ui4, i32], original_type = tensor<4096x32x128xi4>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>) outs(%12 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], original_type = tensor<4096x32xi32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>) {
   ^bb0(%in: i16, %in_2: i4, %out: i32):
     %15 = arith.extsi %in : i16 to i32
     %16 = arith.extui %in_2 : i4 to i32
     %17 = arith.muli %15, %16 : i32
     %18 = arith.addi %17, %out : i32
     linalg.yield %18 : i32
-  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], original_type = tensor<4096x32xi32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
-  %14 = iree_encoding.unset_encoding %13 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i16, ui4, i32], original_type = tensor<4096x32xi32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xi32>
+  } -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], original_type = tensor<4096x32xi32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>>
+  %14 = iree_encoding.unset_encoding %13 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i16, ui4, i32], original_type = tensor<4096x32xi32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xi32>
   %extracted_slice = tensor.extract_slice %14[0, 0] [4096, 32] [1, 1] : tensor<?x?xi32> to tensor<4096x32xi32>
   return %extracted_slice : tensor<4096x32xi32>
 }
diff --git a/compiler/src/iree/compiler/Codegen/Common/CPU/test/lower_to_ukernel_ops.mlir b/compiler/src/iree/compiler/Codegen/Common/CPU/test/lower_to_ukernel_ops.mlir
index d7f30d4..e546c54 100644
--- a/compiler/src/iree/compiler/Codegen/Common/CPU/test/lower_to_ukernel_ops.mlir
+++ b/compiler/src/iree/compiler/Codegen/Common/CPU/test/lower_to_ukernel_ops.mlir
@@ -511,7 +511,7 @@
 func.func @query_tile_sizes_2d() -> (index, index)  attributes {
   hal.executable.target = #hal.executable.target<"vmvx", "vmvx-bytecode-fb", {ukernels = "all"}>
 } {
-  %result:2 = iree_codegen.query_tile_sizes tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %result:2 = iree_codegen.query_tile_sizes tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   return %result#0, %result#1 : index, index
 }
 
diff --git a/compiler/src/iree/compiler/Codegen/Common/CPU/test/vmvx_materialize_encoding.mlir b/compiler/src/iree/compiler/Codegen/Common/CPU/test/vmvx_materialize_encoding.mlir
index e39b0fc..c1bb279 100644
--- a/compiler/src/iree/compiler/Codegen/Common/CPU/test/vmvx_materialize_encoding.mlir
+++ b/compiler/src/iree/compiler/Codegen/Common/CPU/test/vmvx_materialize_encoding.mlir
@@ -11,28 +11,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 
@@ -42,17 +42,17 @@
 //  CHECK-DAG:   %[[M:.+]] = hal.interface.constant.load[0]
 //  CHECK-DAG:   %[[N:.+]] = hal.interface.constant.load[1]
 //  CHECK-DAG:   %[[K:.+]] = hal.interface.constant.load[2]
-//      CHECK:   %[[LHS_TILE_SIZES:.+]]:2 = iree_codegen.query_tile_sizes tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZES:.+]]:2 = iree_codegen.query_tile_sizes tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
 //  CHECK-DAG:   %[[LHS_OUTER_SIZE0:.+]] = affine.apply #[[MAP_CEILDIV]]()[%[[M]], %[[LHS_TILE_SIZES]]#0]
 //  CHECK-DAG:   %[[LHS_OUTER_SIZE1:.+]] = affine.apply #[[MAP_CEILDIV]]()[%[[K]], %[[LHS_TILE_SIZES]]#1]
 //      CHECK:   %[[LHS_BINDING:.+]] = hal.interface.binding.subspan set(0) binding(0)
 // CHECK-SAME:       !flow.dispatch.tensor<readonly:tensor<?x?x?x?xi8>>{%[[LHS_OUTER_SIZE0]], %[[LHS_OUTER_SIZE1]], %[[LHS_TILE_SIZES]]#0, %[[LHS_TILE_SIZES]]#1}
-//      CHECK:   %[[RHS_TILE_SIZES:.+]]:2 = iree_codegen.query_tile_sizes tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1 : i64, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+//      CHECK:   %[[RHS_TILE_SIZES:.+]]:2 = iree_codegen.query_tile_sizes tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1 : i64, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
 //  CHECK-DAG:   %[[RHS_OUTER_SIZE0:.+]] = affine.apply #[[MAP_CEILDIV]]()[%[[N]], %[[RHS_TILE_SIZES]]#0]
 //  CHECK-DAG:   %[[RHS_OUTER_SIZE1:.+]] = affine.apply #[[MAP_CEILDIV]]()[%[[K]], %[[RHS_TILE_SIZES]]#1]
 //      CHECK:   %[[RHS_BINDING:.+]] = hal.interface.binding.subspan set(0) binding(1)
 // CHECK-SAME:       !flow.dispatch.tensor<readonly:tensor<?x?x?x?xi8>>{%[[RHS_OUTER_SIZE0]], %[[RHS_OUTER_SIZE1]], %[[RHS_TILE_SIZES]]#0, %[[RHS_TILE_SIZES]]#1}
-//      CHECK:   %[[RESULT_TILE_SIZES:.+]]:2 = iree_codegen.query_tile_sizes tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2 : i64, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+//      CHECK:   %[[RESULT_TILE_SIZES:.+]]:2 = iree_codegen.query_tile_sizes tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2 : i64, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
 //  CHECK-DAG:   %[[RESULT_OUTER_SIZE0:.+]] = affine.apply #[[MAP_CEILDIV]]()[%[[M]], %[[RESULT_TILE_SIZES]]#0]
 //  CHECK-DAG:   %[[RESULT_OUTER_SIZE1:.+]] = affine.apply #[[MAP_CEILDIV]]()[%[[N]], %[[RESULT_TILE_SIZES]]#1]
 //      CHECK:   %[[OUTS_BINDING:.+]] = hal.interface.binding.subspan set(0) binding(2)
@@ -82,17 +82,17 @@
   %c32_i64 = arith.constant 32 : i64
   %cst = arith.constant 0.000000e+00 : f32
   %c0 = arith.constant 0 : index
-  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<1x2xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg0, %arg1}
-  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<2x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg2, %arg3}
-  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg4, %arg5}
-  %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%arg0, %arg1], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<1x2xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg0, %arg1} -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<1x2xf32>, user_indexing_maps = [#map2, #map3, #map4]>>
-  %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%arg2, %arg3], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<2x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg2, %arg3} -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<2x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>
+  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x2xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg0, %arg1}
+  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg2, %arg3}
+  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg4, %arg5}
+  %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%arg0, %arg1], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x2xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg0, %arg1} -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x2xf32>, user_indexing_maps = [#map2, #map3, #map4]>>
+  %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%arg2, %arg3], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg2, %arg3} -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>
   %5 = affine.apply #map()[%arg6]
   %6 = affine.apply #map1()[%arg7]
-  %7 = tensor.empty(%6, %5) : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>
-  %8 = linalg.fill ins(%cst : f32) outs(%7 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>
-  %9 = linalg.matmul ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<1x2xf32>, user_indexing_maps = [#map2, #map3, #map4]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<2x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>) outs(%8 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>
-  flow.dispatch.tensor.store %9, %2, offsets = [0, 0], sizes = [%arg4, %arg5], strides = [1, 1] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>> -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg4, %arg5}
+  %7 = tensor.empty(%6, %5) : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>
+  %8 = linalg.fill ins(%cst : f32) outs(%7 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>
+  %9 = linalg.matmul ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x2xf32>, user_indexing_maps = [#map2, #map3, #map4]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>) outs(%8 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>
+  flow.dispatch.tensor.store %9, %2, offsets = [0, 0], sizes = [%arg4, %arg5], strides = [1, 1] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>> -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x3xf32>, user_indexing_maps = [#map2, #map3, #map4]>>>{%arg4, %arg5}
   return
 }
 //      CHECK: func.func @fill_matmul
@@ -134,7 +134,7 @@
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
       : !flow.dispatch.tensor<readonly:tensor<?x?xf32>>{%d0, %d1}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%outd0, %outd1}
+      : !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%outd0, %outd1}
   %2 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%d0, %d1], strides = [1, 1]
       : !flow.dispatch.tensor<readonly:tensor<?x?xf32>>{%d0, %d1} -> tensor<?x?xf32>
   %p0 = affine.apply affine_map<()[s0, s1] -> (-s0 + s1)>()[%d0, %outd0]
@@ -143,10 +143,10 @@
   ^bb0(%arg0: index, %arg1: index):
     tensor.yield %cst : f32
   } : tensor<?x?xf32> to tensor<?x?xf32>
-  %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %3, %1, offsets = [0, 0], sizes = [%outd0, %outd1], strides = [1, 1]
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%outd0, %outd1}
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%outd0, %outd1}
   return
 }
 //   CHECK-DAG: #[[MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
@@ -186,14 +186,14 @@
   %outd0 = hal.interface.constant.load [2] : index
   %outd1 = hal.interface.constant.load [3] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%d0, %d1}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%d0, %d1}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
       : !flow.dispatch.tensor<writeonly:tensor<?x?xf32>>{%outd0, %outd1}
   %2 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%d0, %d1], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%d0, %d1}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%d0, %d1}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %3 = iree_encoding.unset_encoding %2
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
   %4 = tensor.extract_slice %3[0, 0] [%outd0, %outd1] [1, 1] : tensor<?x?xf32> to tensor<?x?xf32>
   flow.dispatch.tensor.store %4, %1, offsets = [0, 0], sizes = [%outd0, %outd1], strides = [1, 1]
       : tensor<?x?xf32> -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32>>{%outd0, %outd1}
@@ -232,28 +232,28 @@
   %N = hal.interface.constant.load[1] : index
   %K = hal.interface.constant.load[2] : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
   %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
   %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0)
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%M, %K], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %K}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [%K, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%K, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %5 = flow.dispatch.tensor.load %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      : !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %6 = linalg.matmul
-      ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
-                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      ins(%3, %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>,
+                   tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%5 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   flow.dispatch.tensor.store %6, %2, offsets = [0, 0], sizes = [%M, %N], strides = [1, 1]
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+      -> !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>>{%M, %N}
   return
 }
 //  CHECK-DAG: #[[MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
diff --git a/compiler/src/iree/compiler/Codegen/Common/EncodingUtils.cpp b/compiler/src/iree/compiler/Codegen/Common/EncodingUtils.cpp
index 7eb438b..2f1b3b9 100644
--- a/compiler/src/iree/compiler/Codegen/Common/EncodingUtils.cpp
+++ b/compiler/src/iree/compiler/Codegen/Common/EncodingUtils.cpp
@@ -83,8 +83,10 @@
   auto elemType = tensorType.getElementType();
   OpBuilder builder(context);
 
+  auto opTypeAttr = IREE::Encoding::EncodingOpTypeAttr::get(
+      context, IREE::Encoding::EncodingOpType::matmul);
   auto newEncoding = IREE::Encoding::EncodingAttr::get(
-      context, newIndex, encoding.getElementTypes(),
+      context, newIndex, opTypeAttr, encoding.getElementTypes(),
       TypeAttr::get(RankedTensorType::get(newOriginalShape, elemType)),
       encoding.getMatmulNarrow_N(), encoding.getMatmulNarrow_M(),
       newIndexingMaps, DenseI64ArrayAttr::get(context, newRoundDimsTo));
diff --git a/compiler/src/iree/compiler/Codegen/Common/test/convert_to_destination_passing_style.mlir b/compiler/src/iree/compiler/Codegen/Common/test/convert_to_destination_passing_style.mlir
index eb7e978..68c0873 100644
--- a/compiler/src/iree/compiler/Codegen/Common/test/convert_to_destination_passing_style.mlir
+++ b/compiler/src/iree/compiler/Codegen/Common/test/convert_to_destination_passing_style.mlir
@@ -754,7 +754,7 @@
   %c512 = arith.constant 512 : index
   %c0 = arith.constant 0 : index
   %c16 = arith.constant 16 : index
-  %0:2 = iree_codegen.query_tile_sizes tensor<16x16xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %0:2 = iree_codegen.query_tile_sizes tensor<16x16xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %1 = affine.apply affine_map<()[s0] -> (16 ceildiv s0)>()[%0#0]
   %2 = affine.apply affine_map<()[s0] -> (16 ceildiv s0)>()[%0#1]
   %3 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c512) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<?x?x?x?xi32>>{%1, %2, %0#0, %0#1}
diff --git a/compiler/src/iree/compiler/Codegen/Common/test/materialize_encoding_into_nop.mlir b/compiler/src/iree/compiler/Codegen/Common/test/materialize_encoding_into_nop.mlir
index a527fd1..8e360ee 100644
--- a/compiler/src/iree/compiler/Codegen/Common/test/materialize_encoding_into_nop.mlir
+++ b/compiler/src/iree/compiler/Codegen/Common/test/materialize_encoding_into_nop.mlir
@@ -4,8 +4,8 @@
 #map1 = affine_map<(d0, d1, d2) -> (d2, d1)>
 #map2 = affine_map<(d0, d1, d2) -> (d0, d1)>
 func.func @pack_unpack_gemm_lhs(%arg0 : tensor<?x?xf32>) -> tensor<?x?xf32> {
-  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %1 = iree_encoding.unset_encoding %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
+  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %1 = iree_encoding.unset_encoding %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
   return %1 : tensor<?x?xf32>
 }
 //      CHECK: func @pack_unpack_gemm_lhs(
@@ -23,20 +23,20 @@
     ^bb0(%b0: index, %b1 : index):
       tensor.yield %pad_value : f32
     } : tensor<100x250xf32> to tensor<104x252xf32>
-  %lhs = iree_encoding.set_encoding %pad_lhs : tensor<104x252xf32> -> tensor<104x252xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %lhs = iree_encoding.set_encoding %pad_lhs : tensor<104x252xf32> -> tensor<104x252xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %pad_rhs = tensor.pad %arg1 low[0, 0] high[2, 4] {
     ^bb0(%b0: index, %b1 : index):
       tensor.yield %pad_value : f32
     } : tensor<250x500xf32> to tensor<252x504xf32>
-  %rhs = iree_encoding.set_encoding %pad_rhs : tensor<252x504xf32> -> tensor<252x504xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %rhs = iree_encoding.set_encoding %pad_rhs : tensor<252x504xf32> -> tensor<252x504xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
   %pad_output = tensor.pad %arg2 low[0, 0] high[4, 4] {
     ^bb0(%b0: index, %b1 : index):
       tensor.yield %pad_value : f32
     } : tensor<100x500xf32> to tensor<104x504xf32>
-  %output = iree_encoding.set_encoding %pad_output : tensor<104x504xf32> -> tensor<104x504xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %gemm_packed = linalg.matmul ins(%lhs, %rhs : tensor<104x252xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<252x504xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%output : tensor<104x504xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<104x504xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %gemm = iree_encoding.unset_encoding %gemm_packed : tensor<104x504xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<104x504xf32>
+  %output = iree_encoding.set_encoding %pad_output : tensor<104x504xf32> -> tensor<104x504xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %gemm_packed = linalg.matmul ins(%lhs, %rhs : tensor<104x252xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<252x504xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%output : tensor<104x504xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<104x504xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %gemm = iree_encoding.unset_encoding %gemm_packed : tensor<104x504xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<104x504xf32>
   %result = tensor.extract_slice %gemm[0, 0] [100, 500] [1, 1] : tensor<104x504xf32> to tensor<100x500xf32>
   return %result : tensor<100x500xf32>
 }
@@ -60,12 +60,12 @@
 #map1 = affine_map<(d0, d1, d2) -> (d2, d1)>
 #map2 = affine_map<(d0, d1, d2) -> (d0, d1)>
 func.func @gemm_dynamic(%arg0 : tensor<?x?xf32>, %arg1 : tensor<?x?xf32>, %arg2 : tensor<?x?xf32>) -> tensor<?x?xf32> {
-  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %2 = iree_encoding.set_encoding %arg2 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %3 = linalg.matmul ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
+  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %2 = iree_encoding.set_encoding %arg2 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %3 = linalg.matmul ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
   return %4 : tensor<?x?xf32>
 }
 //      CHECK: func @gemm_dynamic(
@@ -88,14 +88,14 @@
   %cst = arith.constant 0.0 : f32
   %d0 = tensor.dim %arg0, %c0 : tensor<?x?xf32>
   %d1 = tensor.dim %arg1, %c1 : tensor<?x?xf32>
-  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %2 = tensor.empty(%d0, %d1) : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %3 = linalg.fill ins(%cst : f32) outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %4 = linalg.matmul ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %5 = iree_encoding.unset_encoding %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
+  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %2 = tensor.empty(%d0, %d1) : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %3 = linalg.fill ins(%cst : f32) outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %4 = linalg.matmul ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %5 = iree_encoding.unset_encoding %4 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?xf32>
   return %5 : tensor<?x?xf32>
 }
 //      CHECK: func @gemm_fill_dynamic(
@@ -119,12 +119,12 @@
 #map1 = affine_map<(d0, d1, d2, d3) -> (d0, d3, d2)>
 #map2 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2)>
 func.func @batch_matmul(%arg0 : tensor<128x80x32xf32>, %arg1 : tensor<128x32x320xf32>, %arg2 : tensor<128x80x320xf32>) -> tensor<128x80x320xf32> {
-  %0 = iree_encoding.set_encoding %arg0 : tensor<128x80x32xf32> -> tensor<128x80x32xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %1 = iree_encoding.set_encoding %arg1 : tensor<128x32x320xf32> -> tensor<128x32x320xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %2 = iree_encoding.set_encoding %arg2 : tensor<128x80x320xf32> -> tensor<128x80x320xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %3 = linalg.batch_matmul ins(%0, %1 : tensor<128x80x32xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<128x32x320xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%2 : tensor<128x80x320xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<128x80x320xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %4 = iree_encoding.unset_encoding %3 : tensor<128x80x320xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<128x80x320xf32>
+  %0 = iree_encoding.set_encoding %arg0 : tensor<128x80x32xf32> -> tensor<128x80x32xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %1 = iree_encoding.set_encoding %arg1 : tensor<128x32x320xf32> -> tensor<128x32x320xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %2 = iree_encoding.set_encoding %arg2 : tensor<128x80x320xf32> -> tensor<128x80x320xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %3 = linalg.batch_matmul ins(%0, %1 : tensor<128x80x32xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<128x32x320xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%2 : tensor<128x80x320xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<128x80x320xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %4 = iree_encoding.unset_encoding %3 : tensor<128x80x320xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<128x80x320xf32>
   return %4 : tensor<128x80x320xf32>
 }
 //      CHECK: func @batch_matmul(
@@ -142,12 +142,12 @@
 #map1 = affine_map<(d0, d1, d2, d3) -> (d0, d3, d2)>
 #map2 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2)>
 func.func @batch_matmul_dynamic(%arg0 : tensor<?x?x?xf32>, %arg1 : tensor<?x?x?xf32>, %arg2 : tensor<?x?x?xf32>) -> tensor<?x?x?xf32> {
-  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %2 = iree_encoding.set_encoding %arg2 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %3 = linalg.batch_matmul ins(%0, %1 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%2 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %4 = iree_encoding.unset_encoding %3 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?x?xf32>
+  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %2 = iree_encoding.set_encoding %arg2 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %3 = linalg.batch_matmul ins(%0, %1 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%2 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %4 = iree_encoding.unset_encoding %3 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?x?xf32>
   return %4 : tensor<?x?x?xf32>
 }
 //      CHECK: func @batch_matmul_dynamic(
@@ -172,14 +172,14 @@
   %d0 = tensor.dim %arg0, %c0 : tensor<?x?x?xf32>
   %d1 = tensor.dim %arg0, %c1 : tensor<?x?x?xf32>
   %d2 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32>
-  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %2 = tensor.empty(%d0, %d1, %d2) : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %3 = linalg.fill ins(%cst : f32) outs(%2 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %4 = linalg.batch_matmul ins(%0, %1 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
-      outs(%3 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
-  %5 = iree_encoding.unset_encoding %4 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?x?xf32>
+  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %2 = tensor.empty(%d0, %d1, %d2) : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %3 = linalg.fill ins(%cst : f32) outs(%2 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %4 = linalg.batch_matmul ins(%0, %1 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>, tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>)
+      outs(%3 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>) -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>>
+  %5 = iree_encoding.unset_encoding %4 : tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map, #map1, #map2]>> -> tensor<?x?x?xf32>
   return %5 : tensor<?x?x?xf32>
 }
 //      CHECK: func @batch_matmul_fill_dynamic(
@@ -204,10 +204,10 @@
 func.func @drop_encoding_for_hal_flow_ops_static() {
   %c0 = arith.constant 0 : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1x1xf32>>
-  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1x1xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<1x1xf32>, matmul_narrow_M = 1 : index, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>>
+  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1x1xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x1xf32>, matmul_narrow_M = 1 : index, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>>
   %2 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1, 1], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1x1xf32>> -> tensor<1x1xf32>
-  %3 = iree_encoding.set_encoding %2 : tensor<1x1xf32> -> tensor<1x1xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<1x1xf32>, matmul_narrow_M = 1 : index, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
-  flow.dispatch.tensor.store %3, %1, offsets = [0, 0], sizes = [1, 1], strides = [1, 1] : tensor<1x1xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<1x1xf32>, matmul_narrow_M = 1 : index, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>> -> !flow.dispatch.tensor<writeonly:tensor<1x1xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<1x1xf32>, matmul_narrow_M = 1 : index, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>>
+  %3 = iree_encoding.set_encoding %2 : tensor<1x1xf32> -> tensor<1x1xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x1xf32>, matmul_narrow_M = 1 : index, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
+  flow.dispatch.tensor.store %3, %1, offsets = [0, 0], sizes = [1, 1], strides = [1, 1] : tensor<1x1xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x1xf32>, matmul_narrow_M = 1 : index, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>> -> !flow.dispatch.tensor<writeonly:tensor<1x1xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1x1xf32>, matmul_narrow_M = 1 : index, matmul_narrow_N = 1 : index, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>>
   return
 }
 // CHECK-LABEL: func.func @drop_encoding_for_hal_flow_ops_static
@@ -238,10 +238,10 @@
   %14 = flow.dispatch.workload.ordinal %8, 0 : index
   %15 = flow.dispatch.workload.ordinal %13, 1 : index
   %16 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<?x?xbf16>>{%14, %15}
-  %17 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], original_type = tensor<?x?xbf16>, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>>{%14, %15}
+  %17 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<?x?xbf16>, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>>{%14, %15}
   %18 = flow.dispatch.tensor.load %16, offsets = [0, 0], sizes = [%14, %15], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<?x?xbf16>>{%14, %15} -> tensor<?x?xbf16>
-  %19 = iree_encoding.set_encoding %18 : tensor<?x?xbf16> -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], original_type = tensor<?x?xbf16>, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
-  flow.dispatch.tensor.store %19, %17, offsets = [0, 0], sizes = [%14, %15], strides = [1, 1] : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], original_type = tensor<?x?xbf16>, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>> -> !flow.dispatch.tensor<writeonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16], original_type = tensor<?x?xbf16>, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>>{%14, %15}
+  %19 = iree_encoding.set_encoding %18 : tensor<?x?xbf16> -> tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<?x?xbf16>, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>
+  flow.dispatch.tensor.store %19, %17, offsets = [0, 0], sizes = [%14, %15], strides = [1, 1] : tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<?x?xbf16>, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>> -> !flow.dispatch.tensor<writeonly:tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<?x?xbf16>, user_indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>]>>>{%14, %15}
   return
 }
 // CHECK-LABEL: func.func @drop_encoding_for_hal_flow_ops_dynamic
diff --git a/compiler/src/iree/compiler/Codegen/Common/test/tile_and_distribute_to_workgroups.mlir b/compiler/src/iree/compiler/Codegen/Common/test/tile_and_distribute_to_workgroups.mlir
index c9c7d86..02c6886 100644
--- a/compiler/src/iree/compiler/Codegen/Common/test/tile_and_distribute_to_workgroups.mlir
+++ b/compiler/src/iree/compiler/Codegen/Common/test/tile_and_distribute_to_workgroups.mlir
@@ -2232,7 +2232,7 @@
         %cst = arith.constant dense<[-918, -4433, 87, -234, -21393, 7738, 529, -8835, -16817, -375, -199, 572, 5082, 15569, -186, 4955]> : tensor<16xi32>
         %c12544 = arith.constant 12544 : index
         %c16 = arith.constant 16 : index
-        %0:2 = iree_codegen.query_tile_sizes tensor<12544x16xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+        %0:2 = iree_codegen.query_tile_sizes tensor<12544x16xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
         %1 = affine.apply affine_map<()[s0] -> (12544 ceildiv s0)>()[%0#0]
         %2 = affine.apply affine_map<()[s0] -> (16 ceildiv s0)>()[%0#1]
         %3 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c200960) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<?x?x?x?xi32>>{%1, %2, %0#0, %0#1}
diff --git a/compiler/src/iree/compiler/Codegen/VMVX/test/pipeline.mlir b/compiler/src/iree/compiler/Codegen/VMVX/test/pipeline.mlir
index ea0b5c0..6ca9c77 100644
--- a/compiler/src/iree/compiler/Codegen/VMVX/test/pipeline.mlir
+++ b/compiler/src/iree/compiler/Codegen/VMVX/test/pipeline.mlir
@@ -10,15 +10,15 @@
   %c256 = arith.constant 256 : index
   %c512 = arith.constant 512 : index
   %c16 = arith.constant 16 : index
-  %0:2 = iree_codegen.query_tile_sizes tensor<16x16xi8, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %0:2 = iree_codegen.query_tile_sizes tensor<16x16xi8, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %1 = affine.apply #map3()[%0#0]
   %2 = affine.apply #map3()[%0#1]
   %3 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<?x?x?x?xi8>>{%1, %2, %0#0, %0#1}
-  %4:2 = iree_codegen.query_tile_sizes tensor<16x16xi8, #iree_encoding.encoding<operand_index = 1, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %4:2 = iree_codegen.query_tile_sizes tensor<16x16xi8, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %5 = affine.apply #map3()[%4#0]
   %6 = affine.apply #map3()[%4#1]
   %7 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c256) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<?x?x?x?xi8>>{%5, %6, %4#0, %4#1}
-  %8:2 = iree_codegen.query_tile_sizes tensor<16x16xi32, #iree_encoding.encoding<operand_index = 2, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
+  %8:2 = iree_codegen.query_tile_sizes tensor<16x16xi32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#map, #map1, #map2]>> -> index, index
   %9 = affine.apply #map3()[%8#0]
   %10 = affine.apply #map3()[%8#1]
   %11 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c512) : !flow.dispatch.tensor<readwrite:tensor<?x?x?x?xi32>>{%9, %10, %8#0, %8#1}
diff --git a/compiler/src/iree/compiler/Codegen/VMVX/test/select_lowering_strategy.mlir b/compiler/src/iree/compiler/Codegen/VMVX/test/select_lowering_strategy.mlir
index c8eebf4..fc697fc 100644
--- a/compiler/src/iree/compiler/Codegen/VMVX/test/select_lowering_strategy.mlir
+++ b/compiler/src/iree/compiler/Codegen/VMVX/test/select_lowering_strategy.mlir
@@ -168,7 +168,7 @@
   %cst = arith.constant 0.000000e+00 : f32
   %c0 = arith.constant 0 : index
   %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x2048xf32>>
-  %1:2 = iree_codegen.query_tile_sizes tensor<1024x2048xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<1024x2048xf32>>> -> index, index
+  %1:2 = iree_codegen.query_tile_sizes tensor<1024x2048xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1024x2048xf32>>> -> index, index
   %2 = affine.apply #map()[%1#0]
   %3 = affine.apply #map1()[%1#1]
   %4 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<?x?x?x?xf32>>{%2, %3, %1#0, %1#1}
@@ -179,12 +179,12 @@
     %15 = arith.addf %in, %in : f32
     linalg.yield %15 : f32
   } -> tensor<1024x2048xf32>
-  %8:2 = iree_codegen.query_tile_sizes tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<1024x2048xf32>>> -> index, index
+  %8:2 = iree_codegen.query_tile_sizes tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1024x2048xf32>>> -> index, index
   %9 = affine.apply #map()[%8#0]
   %10 = affine.apply #map1()[%8#1]
   %11 = tensor.empty(%9, %10, %8#0, %8#1) : tensor<?x?x?x?xf32>
   %pack = tensor.pack %7 padding_value(%cst : f32) inner_dims_pos = [0, 1] inner_tiles = [%8#0, %8#1] into %11 : tensor<1024x2048xf32> -> tensor<?x?x?x?xf32>
-  %12:2 = iree_codegen.query_tile_sizes tensor<1024x2048xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<1024x2048xf32>>> -> index, index
+  %12:2 = iree_codegen.query_tile_sizes tensor<1024x2048xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<1024x2048xf32>>> -> index, index
   %13 = affine.apply #map()[%12#0]
   %14 = affine.apply #map1()[%12#1]
   flow.dispatch.tensor.store %pack, %4, offsets = [0, 0, 0, 0], sizes = [%13, %14, %12#0, %12#1], strides = [1, 1, 1, 1] : tensor<?x?x?x?xf32> -> !flow.dispatch.tensor<writeonly:tensor<?x?x?x?xf32>>{%13, %14, %12#0, %12#1}
diff --git a/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingBase.td b/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingBase.td
index 9039d08..f829fc8 100644
--- a/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingBase.td
+++ b/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingBase.td
@@ -44,6 +44,19 @@
 class IREEEncoding_EnumAttr<EnumAttrInfo enumInfo, string name = "">
   : EnumAttr<IREEEncoding_Dialect, enumInfo, name>;
 
+// Enums for tagging operand operation in an EncodingAttr
+def MATMUL : I32EnumAttrCase<"matmul", 0>;
+def CONV   : I32EnumAttrCase<"conv", 1>;
+
+def EncodingOpType : IREEEncoding_I32EnumAttr<"EncodingOpType",
+    "Tracks the type of operation of the operand.", [
+      MATMUL,
+      CONV,
+    ]>;
+
+def EncodingOpTypeAttr:
+  IREEEncoding_EnumAttr<EncodingOpType, "optype">;
+
 def EncodingAttr :
     IREEEncoding_Attr<"Encoding"> {
   let mnemonic = "encoding";
@@ -63,6 +76,7 @@
 
   let parameters = (ins
     AttrParameter<"IntegerAttr", "this tensor operand's index in the parameter list">:$operand_index,
+    AttrParameter<"EncodingOpTypeAttr", "operand type">:$op_type,
     AttrParameter<"ArrayAttr", "element types of the user's operands">:$element_types,
     OptionalParameter<"TypeAttr", "type of the original tensor type before padding">:$original_type,
     // TODO(#15466): generalize matmul_narrow_{M,N} into a list?
@@ -75,6 +89,7 @@
 
   let builders = [
     AttrBuilder<(ins "int64_t":$operandIndex,
+        "EncodingOpType":$opType,
         "ArrayRef<Type>":$elemTypes, "Type":$origType,
         CArg<"std::optional<int64_t>", "{}">:$matmulNarrowM,
         CArg<"std::optional<int64_t>", "{}">:$matmulNarrowN,
diff --git a/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingOps.cpp b/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingOps.cpp
index 936c850..d338416 100644
--- a/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingOps.cpp
+++ b/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingOps.cpp
@@ -97,7 +97,8 @@
 //===----------------------------------------------------------------------===//
 
 EncodingAttr EncodingAttr::get(MLIRContext *ctx, int64_t operandIndex,
-                               ArrayRef<Type> elemTypes, Type origType,
+                               EncodingOpType opType, ArrayRef<Type> elemTypes,
+                               Type origType,
                                std::optional<int64_t> matmulNarrowM,
                                std::optional<int64_t> matmulNarrowN,
                                ArrayRef<AffineMap> maps,
@@ -106,14 +107,15 @@
   auto optionalToAttr = [&](std::optional<int64_t> x) {
     return x ? b.getIndexAttr(*x) : IntegerAttr();
   };
+  auto opTypeAttr = EncodingOpTypeAttr::get(ctx, opType);
   auto origTypeAttr = origType ? TypeAttr::get(origType) : TypeAttr();
   auto roundDimsToAttr = roundDimsTo.empty()
                              ? DenseI64ArrayAttr()
                              : b.getDenseI64ArrayAttr(roundDimsTo);
-  return get(ctx, b.getIndexAttr(operandIndex), b.getTypeArrayAttr(elemTypes),
-             origTypeAttr, optionalToAttr(matmulNarrowM),
-             optionalToAttr(matmulNarrowN), b.getAffineMapArrayAttr(maps),
-             roundDimsToAttr);
+  return get(ctx, b.getIndexAttr(operandIndex), opTypeAttr,
+             b.getTypeArrayAttr(elemTypes), origTypeAttr,
+             optionalToAttr(matmulNarrowM), optionalToAttr(matmulNarrowN),
+             b.getAffineMapArrayAttr(maps), roundDimsToAttr);
 }
 
 AffineMap EncodingAttr::getMapForOperandIndex() {
diff --git a/compiler/src/iree/compiler/Dialect/Encoding/IR/test/invalid.mlir b/compiler/src/iree/compiler/Dialect/Encoding/IR/test/invalid.mlir
index b129fb6..b2f23a4 100644
--- a/compiler/src/iree/compiler/Dialect/Encoding/IR/test/invalid.mlir
+++ b/compiler/src/iree/compiler/Dialect/Encoding/IR/test/invalid.mlir
@@ -8,9 +8,9 @@
 
 // -----
 
-func.func @illegal_set_encoding_op_with_source_encoding(%arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>) -> tensor<?x?xf32> {
+func.func @illegal_set_encoding_op_with_source_encoding(%arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>) -> tensor<?x?xf32> {
   // expected-error @+1 {{source of set_encoding op cannot have a tensor encoding}}
-  %0 = iree_encoding.set_encoding %arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
+  %0 = iree_encoding.set_encoding %arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
   return %0 : tensor<?x?xf32>
 }
 
@@ -24,18 +24,18 @@
 
 // -----
 
-func.func @illegal_set_encoding_op_with_rank_change(%arg0 : tensor<?x?xf32>) -> tensor<?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
+func.func @illegal_set_encoding_op_with_rank_change(%arg0 : tensor<?x?xf32>) -> tensor<?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
   // expected-error @+1 {{cannot change the rank of the tensor}}
-  %0 = iree_encoding.set_encoding %arg0: tensor<?x?xf32> -> tensor<?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  return %0 : tensor<?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+  %0 = iree_encoding.set_encoding %arg0: tensor<?x?xf32> -> tensor<?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  return %0 : tensor<?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 
 // -----
 
-func.func @illegal_set_encoding_op_with_shape_change(%arg0 : tensor<10x20xf32>) -> tensor<20x30xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
+func.func @illegal_set_encoding_op_with_shape_change(%arg0 : tensor<10x20xf32>) -> tensor<20x30xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
   // expected-error @+1 {{expected to preserve the logical shape of the tensor}}
-  %0 = iree_encoding.set_encoding %arg0: tensor<10x20xf32> -> tensor<20x30xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  return %0 : tensor<20x30xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+  %0 = iree_encoding.set_encoding %arg0: tensor<10x20xf32> -> tensor<20x30xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  return %0 : tensor<20x30xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 
 // -----
@@ -48,10 +48,10 @@
 
 // -----
 
-func.func @illegal_unset_encoding_op_with_result_encoding(%arg0 : tensor<?x?xf32>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
+func.func @illegal_unset_encoding_op_with_result_encoding(%arg0 : tensor<?x?xf32>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
   // expected-error @+1 {{result of unset_encoding op cannot have a tensor encoding}}
-  %0 = iree_encoding.unset_encoding %arg0: tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  return %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+  %0 = iree_encoding.unset_encoding %arg0: tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  return %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 
 // -----
@@ -64,16 +64,16 @@
 
 // -----
 
-func.func @illegal_unset_encoding_op_with_rank_change(%arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>) -> tensor<?xf32> {
+func.func @illegal_unset_encoding_op_with_rank_change(%arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>) -> tensor<?xf32> {
   // expected-error @+1 {{cannot change the rank of the tensor}}
-  %0 = iree_encoding.unset_encoding %arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> -> tensor<?xf32>
+  %0 = iree_encoding.unset_encoding %arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?xf32>
   return %0 : tensor<?xf32>
 }
 
 // -----
 
-func.func @illegal_unset_encoding_op_with_shape_change(%arg0 : tensor<20x30xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>) -> tensor<10x20xf32> {
+func.func @illegal_unset_encoding_op_with_shape_change(%arg0 : tensor<20x30xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>) -> tensor<10x20xf32> {
   // expected-error @+1 {{expected to preserve the logical shape of the tensor}}
-  %0 = iree_encoding.unset_encoding %arg0: tensor<20x30xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> -> tensor<10x20xf32>
+  %0 = iree_encoding.unset_encoding %arg0: tensor<20x30xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<10x20xf32>
   return %0 : tensor<10x20xf32>
 }
diff --git a/compiler/src/iree/compiler/Dialect/Encoding/IR/test/roundtrip.mlir b/compiler/src/iree/compiler/Dialect/Encoding/IR/test/roundtrip.mlir
index cd6048d..6049b8b 100644
--- a/compiler/src/iree/compiler/Dialect/Encoding/IR/test/roundtrip.mlir
+++ b/compiler/src/iree/compiler/Dialect/Encoding/IR/test/roundtrip.mlir
@@ -1,55 +1,55 @@
 // RUN: iree-opt --split-input-file %s | FileCheck %s
 
 // CHECK: @set_encoding_ops(%[[ARG0:.+]]: tensor<?x?xf32>)
-func.func @set_encoding_ops(%arg0: tensor<?x?xf32>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
-  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
-  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  return %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+func.func @set_encoding_ops(%arg0: tensor<?x?xf32>) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
+  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
+  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  return %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 
 // -----
 
 // CHECK: @set_encoding_ops_mixed_dynamic_static(%[[ARG0:.+]]: tensor<?x10xf32>)
-func.func @set_encoding_ops_mixed_dynamic_static(%arg0: tensor<?x10xf32>) -> tensor<20x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
-  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x10xf32> -> tensor<20x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
-  %0 = iree_encoding.set_encoding %arg0 : tensor<?x10xf32> -> tensor<20x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  return %0 : tensor<20x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+func.func @set_encoding_ops_mixed_dynamic_static(%arg0: tensor<?x10xf32>) -> tensor<20x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
+  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x10xf32> -> tensor<20x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
+  %0 = iree_encoding.set_encoding %arg0 : tensor<?x10xf32> -> tensor<20x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  return %0 : tensor<20x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 
 // -----
 
 // CHECK: @set_encoding_with_batch_matmul_user(%[[ARG0:.+]]: tensor<?x?x?xf32>)
 func.func @set_encoding_with_batch_matmul_user(%arg0: tensor<?x?x?xf32>) {
-  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
-  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [i8, i8, i32]>>
-  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [i8, i8, i32]>>
-  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f16, f16, f32]>>
-  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f32]>>
-  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f16, f16, f16]>>
-  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f16, f16, f16]>>
-  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [bf16, bf16, f32]>>
-  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, f32]>>
-  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [bf16, bf16, bf16]>>
-  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [bf16, bf16, bf16]>>
+  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
+  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [i8, i8, i32]>>
+  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [i8, i8, i32]>>
+  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f16, f16, f32]>>
+  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f32]>>
+  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f16, f16, f16]>>
+  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f16, f16, f16]>>
+  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [bf16, bf16, f32]>>
+  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, f32]>>
+  // CHECK: iree_encoding.set_encoding %[[ARG0]] : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [bf16, bf16, bf16]>>
+  iree_encoding.set_encoding %arg0 : tensor<?x?x?xf32> -> tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [bf16, bf16, bf16]>>
   return
 }
 
 // -----
 
-// CHECK: @unset_encoding_ops(%[[ARG0:.+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, element_types = [f32, f32, f32]>>)
-func.func @unset_encoding_ops(%arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32]>>) -> tensor<?x?xf32> {
-  // CHECK: iree_encoding.unset_encoding %[[ARG0]] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
-  %0 = iree_encoding.unset_encoding %arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
+// CHECK: @unset_encoding_ops(%[[ARG0:.+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, op_type = matmul, element_types = [f32, f32, f32]>>)
+func.func @unset_encoding_ops(%arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32]>>) -> tensor<?x?xf32> {
+  // CHECK: iree_encoding.unset_encoding %[[ARG0]] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
+  %0 = iree_encoding.unset_encoding %arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
   return %0 : tensor<?x?xf32>
 }
 
 // -----
 
-// CHECK: @unset_encoding_ops_mixed_dynamic_static(%[[ARG0:.+]]: tensor<10x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, element_types = [f32, f32, f32]>>)
-func.func @unset_encoding_ops_mixed_dynamic_static(%arg0: tensor<10x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32]>>) -> tensor<?x20xf32> {
-  // CHECK: iree_encoding.unset_encoding %[[ARG0]] : tensor<10x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, element_types = [f32, f32, f32]>>
-  %0 = iree_encoding.unset_encoding %arg0 : tensor<10x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32]>> -> tensor<?x20xf32>
+// CHECK: @unset_encoding_ops_mixed_dynamic_static(%[[ARG0:.+]]: tensor<10x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, op_type = matmul, element_types = [f32, f32, f32]>>)
+func.func @unset_encoding_ops_mixed_dynamic_static(%arg0: tensor<10x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32]>>) -> tensor<?x20xf32> {
+  // CHECK: iree_encoding.unset_encoding %[[ARG0]] : tensor<10x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
+  %0 = iree_encoding.unset_encoding %arg0 : tensor<10x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?x20xf32>
   return %0 : tensor<?x20xf32>
 }
 
@@ -57,14 +57,14 @@
 
 func.func @encoding_tensors_with_ops(%arg0 : tensor<?x?xf32>,
     %arg1 : tensor<?x?xf32>, %arg2 : tensor<?x?xf32>) -> tensor<?x?xf32> {
-  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32]>>
-  %2 = iree_encoding.set_encoding %arg2 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>
+  %0 = iree_encoding.set_encoding %arg0 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  %1 = iree_encoding.set_encoding %arg1 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32]>>
+  %2 = iree_encoding.set_encoding %arg2 : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>
   %3 = linalg.matmul
-      ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32]>>)
-      outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>
-  %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
+      ins(%0, %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>, tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32]>>)
+      outs(%2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>
+  %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
   return %4 : tensor<?x?xf32>
 }
 // CHECK-LABEL: func.func @encoding_tensors_with_ops
@@ -72,11 +72,11 @@
 //  CHECK-SAME:     %[[ARG1:[a-zA-Z0-9]+]]: tensor<?x?xf32>
 //  CHECK-SAME:     %[[ARG2:[a-zA-Z0-9]+]]: tensor<?x?xf32>
 //       CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[ARG0]]
-//  CHECK-SAME:       tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
+//  CHECK-SAME:       tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 //       CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[ARG1]]
-//  CHECK-SAME:       tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, element_types = [f32, f32, f32]>>
+//  CHECK-SAME:       tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 //       CHECK:   %[[OUT:.+]] = iree_encoding.set_encoding %[[ARG2]]
-//  CHECK-SAME:       tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : i64, element_types = [f32, f32, f32]>>
+//  CHECK-SAME:       tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 //       CHECK:   %[[GEMM:.+]] = linalg.matmul
 //  CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 //  CHECK-SAME:       outs(%[[OUT]] :
diff --git a/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/annotate_dispatches.mlir b/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/annotate_dispatches.mlir
index ee3efb7..87f113a 100644
--- a/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/annotate_dispatches.mlir
+++ b/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/annotate_dispatches.mlir
@@ -115,19 +115,19 @@
   // CHECK: flow.executable.export public @dispatch0_map_DxD_f32
   flow.executable.export public @dispatch0
   builtin.module {
-    func.func @dispatch0(%arg0: !flow.dispatch.tensor<readonly:tensor<?x?xf32>>, %arg1: !flow.dispatch.tensor<readonly:tensor<?x?xf32>>, %arg2: index, %arg3: index, %arg4: index, %arg5: index, %arg6: !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>>) {
+    func.func @dispatch0(%arg0: !flow.dispatch.tensor<readonly:tensor<?x?xf32>>, %arg1: !flow.dispatch.tensor<readonly:tensor<?x?xf32>>, %arg2: index, %arg3: index, %arg4: index, %arg5: index, %arg6: !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>>) {
       %0 = flow.dispatch.workload.ordinal %arg2, 0 : index
       %1 = flow.dispatch.workload.ordinal %arg3, 1 : index
       %2 = flow.dispatch.workload.ordinal %arg4, 2 : index
       %3 = flow.dispatch.workload.ordinal %arg5, 3 : index
       %4 = flow.dispatch.tie_shape %arg0 : !flow.dispatch.tensor<readonly:tensor<?x?xf32>>{%0, %1}
       %5 = flow.dispatch.tie_shape %arg1 : !flow.dispatch.tensor<readonly:tensor<?x?xf32>>{%2, %3}
-      %6 = flow.dispatch.tie_shape %arg6 : !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>>{%2, %3}
+      %6 = flow.dispatch.tie_shape %arg6 : !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>>{%2, %3}
       %7 = flow.dispatch.tensor.load %4, offsets = [0, 0], sizes = [%0, %1], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<?x?xf32>>{%0, %1} -> tensor<?x?xf32>
       %8 = flow.dispatch.tensor.load %5, offsets = [0, 0], sizes = [%2, %3], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<?x?xf32>>{%2, %3} -> tensor<?x?xf32>
       %mapped = linalg.map { math.absf } ins(%7 : tensor<?x?xf32>) outs(%8 : tensor<?x?xf32>)
-      %9 = iree_encoding.set_encoding %mapped : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-      flow.dispatch.tensor.store %9, %6, offsets = [0, 0], sizes = [%2, %3], strides = [1, 1] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>>{%arg4, %arg5}
+      %9 = iree_encoding.set_encoding %mapped : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+      flow.dispatch.tensor.store %9, %6, offsets = [0, 0], sizes = [%2, %3], strides = [1, 1] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>>{%arg4, %arg5}
       return
     }
   }
@@ -136,11 +136,11 @@
   // CHECK: flow.executable.export public @dispatch1_unset_encoding_LHS_DxD
   flow.executable.export public @dispatch1
   builtin.module {
-    func.func @dispatch1(%arg0: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>>, %arg1: index, %arg2: index, %arg3: !flow.dispatch.tensor<writeonly:tensor<?x?xf32>>) {
-      %0 = flow.dispatch.tie_shape %arg0 : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>>{%arg1, %arg2}
+    func.func @dispatch1(%arg0: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>>, %arg1: index, %arg2: index, %arg3: !flow.dispatch.tensor<writeonly:tensor<?x?xf32>>) {
+      %0 = flow.dispatch.tie_shape %arg0 : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>>{%arg1, %arg2}
       %1 = flow.dispatch.tie_shape %arg3 : !flow.dispatch.tensor<writeonly:tensor<?x?xf32>>{%arg1, %arg2}
-      %2 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%arg1, %arg2], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>>{%arg1, %arg2} -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-      %3 = iree_encoding.unset_encoding %2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
+      %2 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [%arg1, %arg2], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>>{%arg1, %arg2} -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+      %3 = iree_encoding.unset_encoding %2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
       flow.dispatch.tensor.store %3, %1, offsets = [0, 0], sizes = [%arg1, %arg2], strides = [1, 1] : tensor<?x?xf32> -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32>>{%arg1, %arg2}
       return
     }
diff --git a/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/dispatch_linalg_on_tensors.mlir b/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/dispatch_linalg_on_tensors.mlir
index 2741edb..8f14a27 100644
--- a/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/dispatch_linalg_on_tensors.mlir
+++ b/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/dispatch_linalg_on_tensors.mlir
@@ -1875,10 +1875,10 @@
 // -----
 
 util.func public @set_encoding_op(%arg0 : tensor<?x?xf32>)
-    -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
+    -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
   %0 = iree_encoding.set_encoding %arg0
-      : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  util.return %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+      : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  util.return %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 //      CHECK: util.func public @set_encoding_op
 // CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xf32>
@@ -1890,7 +1890,7 @@
 // CHECK-NEXT:     %[[INARG:.+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32>>
 // CHECK-SAME:     %[[INDEXARG0:[a-zA-Z0-9]+]]: index
 // CHECK-SAME:     %[[INDEXARG1:[a-zA-Z0-9]+]]: index
-// CHECK-SAME:     %[[OUTARG:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>
+// CHECK-SAME:     %[[OUTARG:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
 //  CHECK-DAG:     %[[W0:.+]] = flow.dispatch.workload.ordinal %[[INDEXARG0]], 0
 //  CHECK-DAG:     %[[W1:.+]] = flow.dispatch.workload.ordinal %[[INDEXARG1]], 1
 //      CHECK:     %[[LOAD:.+]] = flow.dispatch.tensor.load %[[INARG]]
@@ -1898,7 +1898,7 @@
 //      CHECK:     %[[ENCODING:.+]] = iree_encoding.set_encoding %[[LOAD]]
 //      CHECK:     flow.dispatch.tensor.store %[[ENCODING]], %[[OUTARG]]
 // CHECK-SAME:         sizes = [%[[W0]], %[[W1]]]
-// CHECK-SAME:         !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>{%[[W0]], %[[W1]]}
+// CHECK-SAME:         !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>{%[[W0]], %[[W1]]}
 //      CHECK:     flow.return
 //      CHECK:   count(%[[WL0:[a-zA-Z0-9]+]]: index, %[[WL1:[a-zA-Z0-9]+]]: index)
 //      CHECK:     %[[X:[a-zA-Z0-9]+]], %[[Y:[a-zA-Z0-9]+]], %[[Z:.+]] = flow.dispatch.workgroup_count_from_slice %[[WL0]], %[[WL1]]
@@ -1907,20 +1907,20 @@
 
 // -----
 
-util.func public @unset_encoding_op(%arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>)
+util.func public @unset_encoding_op(%arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>)
     -> tensor<?x?xf32> {
   %0 = iree_encoding.unset_encoding %arg0
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
   util.return %0 : tensor<?x?xf32>
 }
 //      CHECK: util.func public @unset_encoding_op
-// CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
+// CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 //  CHECK-DAG:   %[[C0:.+]] = arith.constant 0 : index
 //  CHECK-DAG:   %[[C1:.+]] = arith.constant 1 : index
 //  CHECK-DAG:   %[[D0:.+]] = tensor.dim %[[ARG0]], %[[C0]]
 //  CHECK-DAG:   %[[D1:.+]] = tensor.dim %[[ARG0]], %[[C1]]
 //      CHECK:   %[[DISPATCH:.+]] = flow.dispatch.workgroups[%[[D0]], %[[D1]]](%[[ARG0]], %[[D0]], %[[D1]])
-// CHECK-NEXT:       %[[INARG:.+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>
+// CHECK-NEXT:       %[[INARG:.+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
 // CHECK-SAME:       %[[INDEXARG0:[a-zA-Z0-9]+]]: index
 // CHECK-SAME:       %[[INDEXARG1:[a-zA-Z0-9]+]]: index
 // CHECK-SAME:       %[[OUTARG:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<writeonly:tensor<?x?xf32>>
@@ -1928,7 +1928,7 @@
 //  CHECK-DAG:     %[[D1_W:.+]] = flow.dispatch.workload.ordinal %[[INDEXARG1]], 1
 //      CHECK:     %[[LOAD:.+]] = flow.dispatch.tensor.load %[[INARG]]
 // CHECK-SAME:         sizes = [%[[D0_W]], %[[D1_W]]]
-// CHECK-SAME:         !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>{%[[D0_W]], %[[D1_W]]}
+// CHECK-SAME:         !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>{%[[D0_W]], %[[D1_W]]}
 //      CHECK:     %[[ENCODING:.+]] = iree_encoding.unset_encoding %[[LOAD]]
 //      CHECK:     flow.dispatch.tensor.store %[[ENCODING]], %[[OUTARG]]
 // CHECK-SAME:         !flow.dispatch.tensor<writeonly:tensor<?x?xf32>>{%[[D0_W]], %[[D1_W]]}
@@ -1942,7 +1942,7 @@
 
 #map = affine_map<()[s0] -> (-s0 + (s0 ceildiv 16) * 16)>
 util.func public @pad_and_set_encoding_op(%arg0 : tensor<?x?xf32>)
-    -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
+    -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
   %c0 = arith.constant 0 : index
   %c1 = arith.constant 1 : index
   %cst = arith.constant 0.0 : f32
@@ -1955,8 +1955,8 @@
       tensor.yield %cst : f32
     } : tensor<?x?xf32> to tensor<?x?xf32>
   %encoding = iree_encoding.set_encoding %pad
-      : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  util.return %encoding : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+      : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  util.return %encoding : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 //  CHECK-DAG: #[[MAP0:.+]] = affine_map<()[s0] -> ((s0 ceildiv 16) * 16)>
 //  CHECK-DAG: #[[MAP1:.+]] = affine_map<()[s0] -> (-s0 + (s0 ceildiv 16) * 16)>
@@ -1974,7 +1974,7 @@
 // CHECK-SAME:       %[[INARG:.+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32>>
 // CHECK-SAME:       %[[PADDED_D0:[a-zA-Z0-9]+]]: index
 // CHECK-SAME:       %[[PADDED_D1:[a-zA-Z0-9]+]]: index
-// CHECK-SAME:       %[[OUTARG:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>
+// CHECK-SAME:       %[[OUTARG:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
 //  CHECK-DAG:     %[[D1_W:.+]] = flow.dispatch.workload.ordinal %[[INDEXARG1]], 0
 //  CHECK-DAG:     %[[D0_W:.+]] = flow.dispatch.workload.ordinal %[[INDEXARG0]], 1
 //  CHECK-DAG:     %[[PADDED_D0_W:.+]] = flow.dispatch.workload.ordinal %[[PADDED_D0]], 2
@@ -1987,7 +1987,7 @@
 //      CHECK:     %[[SET_ENCODING:.+]] = iree_encoding.set_encoding %[[PADDED]]
 //      CHECK:     flow.dispatch.tensor.store %[[SET_ENCODING]], %[[OUTARG]]
 // CHECK-SAME:         sizes = [%[[PADDED_D0_W]], %[[PADDED_D1_W]]]
-// CHECK-SAME:         !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>{%[[PADDED_D0_W]], %[[PADDED_D1_W]]}
+// CHECK-SAME:         !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>{%[[PADDED_D0_W]], %[[PADDED_D1_W]]}
 //      CHECK:     flow.return
 //      CHECK:   count(%[[WL0:[a-zA-Z0-9]+]]: index, %[[WL1:[a-zA-Z0-9]+]]: index,
 // CHECK-SAME:         %[[WL2:[a-zA-Z0-9]+]]: index, %[[WL3:[a-zA-Z0-9]+]]: index)
@@ -1998,16 +1998,16 @@
 // -----
 
 util.func public @unset_encoding_and_slice(
-    %arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>,
+    %arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>,
     %arg1 : index, %arg2 : index) -> tensor<?x?xf32> {
   %0 = iree_encoding.unset_encoding %arg0
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
   %1 = tensor.extract_slice %0[0, 0] [%arg1, %arg2] [1, 1]
       : tensor<?x?xf32> to tensor<?x?xf32>
   util.return %1 : tensor<?x?xf32>
 }
 //      CHECK: util.func public @unset_encoding_and_slice
-// CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
+// CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 // CHECK-SAME:     %[[ARG1:[a-zA-Z0-9]+]]: index
 // CHECK-SAME:     %[[ARG2:[a-zA-Z0-9]+]]: index
 //  CHECK-DAG:   %[[C0:.+]] = arith.constant 0 : index
@@ -2015,7 +2015,7 @@
 //  CHECK-DAG:   %[[D0:.+]] = tensor.dim %[[ARG0]], %[[C0]]
 //  CHECK-DAG:   %[[D1:.+]] = tensor.dim %[[ARG0]], %[[C1]]
 //      CHECK:   %[[DISPATCH:.+]] = flow.dispatch.workgroups[%[[D0]], %[[D1]], %[[ARG1]], %[[ARG2]]](%[[ARG0]], %[[D0]], %[[D1]], %[[ARG1]], %[[ARG2]])
-// CHECK-NEXT:       %[[INARG:.+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>
+// CHECK-NEXT:       %[[INARG:.+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
 // CHECK-SAME:       %[[INDEXARG0:[a-zA-Z0-9]+]]: index
 // CHECK-SAME:       %[[INDEXARG1:[a-zA-Z0-9]+]]: index
 // CHECK-SAME:       %[[INDEXARG2:[a-zA-Z0-9]+]]: index
@@ -2027,7 +2027,7 @@
 //  CHECK-DAG:     %[[ARG1_W:.+]] = flow.dispatch.workload.ordinal %[[INDEXARG3]], 3
 //      CHECK:     %[[LOAD:.+]] = flow.dispatch.tensor.load %[[INARG]]
 // CHECK-SAME:         sizes = [%[[D0_W]], %[[D1_W]]]
-// CHECK-SAME:         !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>{%[[D0_W]], %[[D1_W]]}
+// CHECK-SAME:         !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>{%[[D0_W]], %[[D1_W]]}
 //      CHECK:     %[[ENCODING:.+]] = iree_encoding.unset_encoding %[[LOAD]]
 //      CHECK:     %[[SLICE:.+]] = tensor.extract_slice %[[ENCODING]][0, 0] [%[[ARG0_W]], %[[ARG1_W]]]
 //      CHECK:     flow.dispatch.tensor.store %[[SLICE]], %[[OUTARG]]
@@ -2040,9 +2040,9 @@
 #map = affine_map<(d0, d1) -> (d1)>
 #map1 = affine_map<(d0, d1) -> (d0, d1)>
 util.func public @root_on_unset_encoding(
-    %arg0: tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>,
+    %arg0: tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>,
     %arg1: tensor<96xf32>) -> tensor<784x96xf32> {
-  %0 = iree_encoding.unset_encoding %arg0 : tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> -> tensor<784x96xf32>
+  %0 = iree_encoding.unset_encoding %arg0 : tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<784x96xf32>
   %1 = tensor.empty() : tensor<784x96xf32>
   %cst = arith.constant 0.000000e+00 : f32
   %2 = linalg.fill ins(%cst : f32) outs(%1 : tensor<784x96xf32>) -> tensor<784x96xf32>
@@ -2061,18 +2061,18 @@
 //      CHECL: #[[MAP0:.+]] = affine_map<(d0, d1) -> (d1)>
 //      CHECK: #[[MAP1:.+]] = affine_map<(d0, d1) -> (d0, d1)>
 //      CHECK: util.func public @root_on_unset_encoding
-// CHECK-SAME:     %[[ARG0:.+]]: tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
+// CHECK-SAME:     %[[ARG0:.+]]: tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 // CHECK-SAME:     %[[ARG1:.+]]: tensor<96xf32>
 //      CHECK:   %[[DISPATCH:.+]] = flow.dispatch.workgroups(%[[ARG0]], %[[ARG1]])
-// CHECK-NEXT:     %[[INARG0:.+]]: !flow.dispatch.tensor<readonly:tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>
+// CHECK-NEXT:     %[[INARG0:.+]]: !flow.dispatch.tensor<readonly:tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
 // CHECK-SAME:     %[[INARG1:.+]]: !flow.dispatch.tensor<readonly:tensor<96xf32>>
 // CHECK-SAME:     %[[OUTARG:.+]]: !flow.dispatch.tensor<writeonly:tensor<784x96xf32>>
 //      CHECK:     %[[LOAD0:.+]] = flow.dispatch.tensor.load %[[INARG0]], offsets = [0, 0], sizes = [784, 96], strides = [1, 1]
-// CHECK-SAME:       : !flow.dispatch.tensor<readonly:tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>
+// CHECK-SAME:       : !flow.dispatch.tensor<readonly:tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
 //      CHECK:     %[[LOAD1:.+]] = flow.dispatch.tensor.load %[[INARG1]], offsets = [0], sizes = [96], strides = [1]
 // CHECK-SAME:       : !flow.dispatch.tensor<readonly:tensor<96xf32>>
 //      CHECK:     %[[OUT:.+]] = tensor.empty() : tensor<784x96xf32>
-//      CHECK:     %[[UNSET:.+]] = iree_encoding.unset_encoding %[[LOAD0]] : tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>> -> tensor<784x96xf32>
+//      CHECK:     %[[UNSET:.+]] = iree_encoding.unset_encoding %[[LOAD0]] : tensor<784x96xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<784x96xf32>
 //      CHECK:     %[[GENERIC0:.+]] = linalg.generic {indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP1]]],
 // CHECK-SAME:       iterator_types = ["parallel", "parallel"]} ins(%[[LOAD1]], %[[UNSET]]
 // CHECK-SAME:       : tensor<96xf32>, tensor<784x96xf32>) outs(%[[OUT]] : tensor<784x96xf32>)
@@ -2085,26 +2085,26 @@
 // -----
 
 util.func public @gemm_encoded(
-    %arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>,
-    %arg1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32]>>,
-    %arg2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>)
-    -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>> {
+    %arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>,
+    %arg1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32]>>,
+    %arg2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>)
+    -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>> {
   %0 = linalg.matmul
       ins(%arg0, %arg1
-          : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>,
-            tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32]>>)
-      outs(%arg2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>
-  util.return %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>
+          : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>,
+            tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32]>>)
+      outs(%arg2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>
+  util.return %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 //      CHECK: util.func public @gemm_encoded
-// CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
-// CHECK-SAME:     %[[ARG1:[a-zA-Z0-9]+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, element_types = [f32, f32, f32]>>
-// CHECK-SAME:     %[[ARG2:[a-zA-Z0-9]+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : i64, element_types = [f32, f32, f32]>>
+// CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
+// CHECK-SAME:     %[[ARG1:[a-zA-Z0-9]+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
+// CHECK-SAME:     %[[ARG2:[a-zA-Z0-9]+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 //      CHECK:   %[[DISPATCH:.+]] = flow.dispatch.workgroups
-// CHECK-NEXT:     %[[LHS_IN:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>
-// CHECK-SAME:     %[[RHS_IN:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, element_types = [f32, f32, f32]>>>
-// CHECK-SAME:     %[[INIT_IN:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : i64, element_types = [f32, f32, f32]>>>
+// CHECK-NEXT:     %[[LHS_IN:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
+// CHECK-SAME:     %[[RHS_IN:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
+// CHECK-SAME:     %[[INIT_IN:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<readwrite:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
 //  CHECK-DAG:     %[[LHS:.+]] = flow.dispatch.tensor.load %[[LHS_IN]]
 //  CHECK-DAG:     %[[RHS:.+]] = flow.dispatch.tensor.load %[[RHS_IN]]
 //  CHECK-DAG:     %[[INIT:.+]] = flow.dispatch.tensor.load %[[INIT_IN]]
@@ -2116,32 +2116,32 @@
 // -----
 
 util.func public @gemm_fill_encoded(
-    %arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>,
-    %arg1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32]>>)
-    -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>> {
+    %arg0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>,
+    %arg1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32]>>)
+    -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>> {
   %c0 = arith.constant 0 : index
   %c1 = arith.constant 1 : index
   %cst = arith.constant 0.0 : f32
-  %d0 = tensor.dim %arg0, %c0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  %d1 = tensor.dim %arg1, %c1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32]>>
-  %empty = tensor.empty(%d0, %d1) : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>
-  %fill = linalg.fill ins(%cst : f32) outs(%empty : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>
+  %d0 = tensor.dim %arg0, %c0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  %d1 = tensor.dim %arg1, %c1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32]>>
+  %empty = tensor.empty(%d0, %d1) : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>
+  %fill = linalg.fill ins(%cst : f32) outs(%empty : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>
   %0 = linalg.matmul
       ins(%arg0, %arg1
-          : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>,
-            tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32]>>)
-      outs(%fill : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>)
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>
-  util.return %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>
+          : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>,
+            tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32]>>)
+      outs(%fill : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>)
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>
+  util.return %0 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 //      CHECK: util.func public @gemm_fill_encoded
-// CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
-// CHECK-SAME:     %[[ARG1:[a-zA-Z0-9]+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, element_types = [f32, f32, f32]>>
+// CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
+// CHECK-SAME:     %[[ARG1:[a-zA-Z0-9]+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 //      CHECK:   %[[DISPATCH:.+]] = flow.dispatch.workgroups
-// CHECK-NEXT:     %[[LHS_IN:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>>
-// CHECK-SAME:     %[[RHS_IN:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, element_types = [f32, f32, f32]>>>
-// CHECK-SAME:     %[[RESULT:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : i64, element_types = [f32, f32, f32]>>>
+// CHECK-NEXT:     %[[LHS_IN:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
+// CHECK-SAME:     %[[RHS_IN:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
+// CHECK-SAME:     %[[RESULT:[a-zA-Z0-9]+]]: !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : i64, op_type = matmul, element_types = [f32, f32, f32]>>>
 //  CHECK-DAG:     %[[LHS:.+]] = flow.dispatch.tensor.load %[[LHS_IN]]
 //  CHECK-DAG:     %[[RHS:.+]] = flow.dispatch.tensor.load %[[RHS_IN]]
 //      CHECK:     %[[EMPTY:.+]] = tensor.empty
diff --git a/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/dispatch_linalg_on_tensors_default.mlir b/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/dispatch_linalg_on_tensors_default.mlir
index 90aceda..17bcc95 100644
--- a/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/dispatch_linalg_on_tensors_default.mlir
+++ b/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/dispatch_linalg_on_tensors_default.mlir
@@ -33,7 +33,7 @@
 #map = affine_map<(d0, d1) -> (d1)>
 #map1 = affine_map<(d0, d1) -> (d0, d1)>
 util.func public @elem_set_encoding(%arg0: tensor<512xf32>, %arg1: tensor<384x512xf32>,
-    %arg2: tensor<384x512xf32>) -> tensor<384x512xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
+    %arg2: tensor<384x512xf32>) -> tensor<384x512xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
   %0 = tensor.empty() : tensor<384x512xf32>
   %1 = linalg.generic {indexing_maps = [#map, #map1, #map1, #map1],
                        iterator_types = ["parallel", "parallel"]}
@@ -44,8 +44,8 @@
     %4 = arith.addf %3, %in_1 : f32
     linalg.yield %4 : f32
   } -> tensor<384x512xf32>
-  %2 = iree_encoding.set_encoding %1 : tensor<384x512xf32> -> tensor<384x512xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  util.return %2 : tensor<384x512xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+  %2 = iree_encoding.set_encoding %1 : tensor<384x512xf32> -> tensor<384x512xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  util.return %2 : tensor<384x512xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 // CHECK-LABEL: util.func public @elem_set_encoding
 // CHECK:         flow.dispatch.workgroups
diff --git a/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/form_dispatch_regions.mlir b/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/form_dispatch_regions.mlir
index ba231b4..022315c 100644
--- a/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/form_dispatch_regions.mlir
+++ b/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/form_dispatch_regions.mlir
@@ -138,7 +138,7 @@
 // -----
 
 util.func public @set_encoding_fusion(%arg0 : tensor<?x?xf32>, %arg1 : tensor<?x?xf32>,
-    %arg2 : index, %arg3 : index) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
+    %arg2 : index, %arg3 : index) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
   %cst = arith.constant 0.0 : f32
   %c0 = arith.constant 0 : index
   %c1 = arith.constant 1 : index
@@ -168,8 +168,8 @@
       linalg.yield %6 : f32
   } -> tensor<?x?xf32>
   %6 = iree_encoding.set_encoding %5
-      : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  util.return %6 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+      : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  util.return %6 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 // CHECK-LABEL: util.func public @set_encoding_fusion(
 //  CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]: tensor<?x?xf32>
@@ -188,15 +188,15 @@
 // -----
 
 util.func public @set_encoding_pad_fusion(%arg0 : tensor<?x?xf32>,
-    %arg1 : index, %arg2 : index) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
+    %arg1 : index, %arg2 : index) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
   %cst = arith.constant 0.0 : f32
   %0 = tensor.pad %arg0 low[0, 0] high[%arg1, %arg2] {
     ^bb0(%b0: index, %b1 : index):
       tensor.yield %cst : f32
   } : tensor<?x?xf32> to tensor<?x?xf32>
   %1 = iree_encoding.set_encoding %0
-      : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  util.return %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+      : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  util.return %1 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 // CHECK-LABEL: util.func public @set_encoding_pad_fusion(
 //  CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xf32>
@@ -209,7 +209,7 @@
 // -----
 
 util.func public @set_encoding_pad_elementwise_fusion(%arg0 : tensor<?x?xf32>, %arg1 : tensor<?x?xf32>,
-    %arg2 : index, %arg3 : index) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
+    %arg2 : index, %arg3 : index) -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
   %cst = arith.constant 0.0 : f32
   %c0 = arith.constant 0 : index
   %c1 = arith.constant 1 : index
@@ -243,8 +243,8 @@
       tensor.yield %cst : f32
   } : tensor<?x?xf32> to tensor<?x?xf32>
   %7 = iree_encoding.set_encoding %6
-      : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  util.return %7 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+      : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  util.return %7 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 // CHECK-LABEL: util.func public @set_encoding_pad_elementwise_fusion(
 //  CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]: tensor<?x?xf32>
@@ -264,12 +264,12 @@
 // -----
 
 util.func public @unset_encoding_elementwise_fusion(
-    %arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>,
+    %arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>,
     %arg1: tensor<?xf32>) -> tensor<?x?xf32> {
   %c0 = arith.constant 0 : index
   %c1 = arith.constant 1 : index
   %0 = iree_encoding.unset_encoding %arg0
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
   %1 = tensor.dim %0, %c0 : tensor<?x?xf32>
   %2 = tensor.dim %0, %c1 : tensor<?x?xf32>
   %3 = tensor.empty(%1, %2) : tensor<?x?xf32>
@@ -287,7 +287,7 @@
   util.return %4 : tensor<?x?xf32>
 }
 // CHECK-LABEL: util.func public @unset_encoding_elementwise_fusion(
-//  CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
+//  CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 //  CHECK-SAME:     %[[ARG1:.+]]: tensor<?xf32>)
 //       CHECK:   %[[RESULT:.+]] = flow.dispatch.region
 //       CHECK:     %[[UNSET_ENCODING:.+]] = iree_encoding.unset_encoding %[[ARG0]]
@@ -299,12 +299,12 @@
 // -----
 
 util.func public @unset_encoding_slice_elementwise_fusion(
-    %arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>,
+    %arg0: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>,
     %arg1: tensor<?xf32>, %arg2 : index, %arg3 : index) -> tensor<?x?xf32> {
   %c0 = arith.constant 0 : index
   %c1 = arith.constant 1 : index
   %0 = iree_encoding.unset_encoding %arg0
-      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
+      : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>
   %1 = tensor.extract_slice %0[0, 0] [%arg2, %arg3] [1, 1] : tensor<?x?xf32> to tensor<?x?xf32>
   %2 = tensor.dim %1, %c0 : tensor<?x?xf32>
   %3 = tensor.dim %1, %c1 : tensor<?x?xf32>
@@ -323,7 +323,7 @@
   util.return %5 : tensor<?x?xf32>
 }
 // CHECK-LABEL: util.func public @unset_encoding_slice_elementwise_fusion(
-//  CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
+//  CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 //  CHECK-SAME:     %[[ARG1:.+]]: tensor<?xf32>
 //       CHECK:   %[[RESULT0:.+]] = flow.dispatch.region
 //       CHECK:     %[[UNSET_ENCODING:.+]] = iree_encoding.unset_encoding %[[ARG0]]
diff --git a/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/tensor_pad_to_tensor_insert_slice.mlir b/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/tensor_pad_to_tensor_insert_slice.mlir
index c9e4d5d..3505806 100644
--- a/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/tensor_pad_to_tensor_insert_slice.mlir
+++ b/compiler/src/iree/compiler/Dialect/Flow/Transforms/test/tensor_pad_to_tensor_insert_slice.mlir
@@ -82,11 +82,11 @@
 util.func public @dispatch_dispatch_0_generic_512x1024_f32(
     %arg0: !flow.dispatch.tensor<readonly:tensor<512x1024xf32>>,
     %arg1: index, %arg2: index, %arg3: index, %arg4: index,
-    %arg5: !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<512x1024xf32>>>>) {
+    %arg5: !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<512x1024xf32>>>>) {
   %cst = arith.constant 0.000000e+00 : f32
   %0 = flow.dispatch.workload.ordinal %arg3, 2 : index
   %1 = flow.dispatch.workload.ordinal %arg4, 3 : index
-  %2 = flow.dispatch.tie_shape %arg5 : !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<512x1024xf32>>>>{%0, %1}
+  %2 = flow.dispatch.tie_shape %arg5 : !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<512x1024xf32>>>>{%0, %1}
   %3 = flow.dispatch.workload.ordinal %arg1, 0 : index
   %4 = flow.dispatch.workload.ordinal %arg2, 1 : index
   %5 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [512, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<512x1024xf32>> -> tensor<512x1024xf32>
@@ -94,8 +94,8 @@
   ^bb0(%arg6: index, %arg7: index):
     tensor.yield %cst : f32
   } : tensor<512x1024xf32> to tensor<?x?xf32>
-  %11 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<512x1024xf32>>>
-  flow.dispatch.tensor.store %11, %2, offsets = [0, 0], sizes = [%0, %1], strides = [1, 1] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<512x1024xf32>>> -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<512x1024xf32>>>>{%0, %1}
+  %11 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<512x1024xf32>>>
+  flow.dispatch.tensor.store %11, %2, offsets = [0, 0], sizes = [%0, %1], strides = [1, 1] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<512x1024xf32>>> -> !flow.dispatch.tensor<writeonly:tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<512x1024xf32>>>>{%0, %1}
   util.return
 }
 
diff --git a/compiler/src/iree/compiler/Dialect/Stream/Transforms/test/encode_host_tensors.mlir b/compiler/src/iree/compiler/Dialect/Stream/Transforms/test/encode_host_tensors.mlir
index 87ae134..765c1d3 100644
--- a/compiler/src/iree/compiler/Dialect/Stream/Transforms/test/encode_host_tensors.mlir
+++ b/compiler/src/iree/compiler/Dialect/Stream/Transforms/test/encode_host_tensors.mlir
@@ -25,7 +25,7 @@
 #map1 = affine_map<(d0, d1, d2) -> (d2, d1)>
 #map2 = affine_map<(d0, d1, d2) -> (d0, d1)>
 util.func public @sizeof_lhs_encoding_dynamic(%arg0: index, %arg1: index) -> index {
-  %0 = stream.tensor.sizeof tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], original_type = tensor<?x?xf32>, user_indexing_maps = [#map, #map1, #map2], round_dims_to = array<i64: 4, 8, 16>>>{%arg0, %arg1} : index
+  %0 = stream.tensor.sizeof tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<?x?xf32>, user_indexing_maps = [#map, #map1, #map2], round_dims_to = array<i64: 4, 8, 16>>>{%arg0, %arg1} : index
   util.return %0 : index
 }
 // CHECK-LABEL: @sizeof_lhs_encoding_dynamic
@@ -45,7 +45,7 @@
 #map1 = affine_map<(d0, d1, d2) -> (d2, d1)>
 #map2 = affine_map<(d0, d1, d2) -> (d0, d1)>
 util.func public @sizeof_rhs_encoding_dynamic(%arg0: index, %arg1: index) -> index {
-  %0 = stream.tensor.sizeof tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, element_types = [f32, f32, f32], original_type = tensor<?x?xf32>, user_indexing_maps = [#map, #map1, #map2], round_dims_to = array<i64: 4, 8, 16>>>{%arg0, %arg1} : index
+  %0 = stream.tensor.sizeof tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<?x?xf32>, user_indexing_maps = [#map, #map1, #map2], round_dims_to = array<i64: 4, 8, 16>>>{%arg0, %arg1} : index
   util.return %0 : index
 }
 // CHECK-LABEL: @sizeof_rhs_encoding_dynamic
@@ -66,7 +66,7 @@
 #map1 = affine_map<(d0, d1, d2) -> (d2, d1)>
 #map2 = affine_map<(d0, d1, d2) -> (d0, d1)>
 util.func public @sizeof_result_encoding_dynamic(%arg0: index, %arg1: index) -> index {
-  %0 = stream.tensor.sizeof tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32], original_type = tensor<?x?xf32>, user_indexing_maps = [#map, #map1, #map2], round_dims_to = array<i64: 4, 8, 16>>>{%arg0, %arg1} : index
+  %0 = stream.tensor.sizeof tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<?x?xf32>, user_indexing_maps = [#map, #map1, #map2], round_dims_to = array<i64: 4, 8, 16>>>{%arg0, %arg1} : index
   util.return %0 : index
 }
 // CHECK-LABEL: @sizeof_result_encoding_dynamic
diff --git a/compiler/src/iree/compiler/GlobalOptimization/SetEncoding.cpp b/compiler/src/iree/compiler/GlobalOptimization/SetEncoding.cpp
index 43114cd..a7f84e6 100644
--- a/compiler/src/iree/compiler/GlobalOptimization/SetEncoding.cpp
+++ b/compiler/src/iree/compiler/GlobalOptimization/SetEncoding.cpp
@@ -145,14 +145,15 @@
                                int64_t operandIndex,
                                ArrayRef<Type> operandElemTypes,
                                MatmulNarrowSizes narrow,
-                               ArrayRef<AffineMap> indexingMaps) {
+                               ArrayRef<AffineMap> indexingMaps,
+                               IREE::Encoding::EncodingOpType opType) {
   MLIRContext *ctx = builder.getContext();
   // No need to specify original_type in the encoding poadded to pad(), because
   // the operand there is the `source` tensor, so it will default to reading its
   // original shape.
-  auto encodingForPad = EncodingAttr::get(ctx, operandIndex, operandElemTypes,
-                                          /*originalType=*/Type{}, narrow.M,
-                                          narrow.N, indexingMaps);
+  auto encodingForPad = EncodingAttr::get(
+      ctx, operandIndex, opType, operandElemTypes,
+      /*originalType=*/Type{}, narrow.M, narrow.N, indexingMaps);
   Value padded = pad(builder, loc, source, encodingForPad);
   // For setEncoding() below, we potentially need to specify an encoding with an
   // explicit original_type, because the operand there is the padded tensor
@@ -163,7 +164,7 @@
   auto encodingForSetEncoding = encodingForPad;
   if (padded.getType() != source.getType()) {
     encodingForSetEncoding = EncodingAttr::get(
-        ctx, operandIndex, operandElemTypes,
+        ctx, operandIndex, opType, operandElemTypes,
         /*originalType=*/source.getType(), narrow.M, narrow.N, indexingMaps);
   }
   return setEncoding(builder, loc, padded, encodingForSetEncoding);
@@ -337,22 +338,23 @@
     SmallVector<AffineMap> maps = linalgOp.getIndexingMapsArray();
     Value encodedLhs, encodedRhs, encodedOut;
 
+    auto opType = IREE::Encoding::EncodingOpType::matmul;
     if (!padFactor) {
       encodedLhs =
           padAndSetEncoding(rewriter, loc, lhs, IREE::Encoding::MATMUL_LHS,
-                            elemTypes, narrowSizes, maps);
+                            elemTypes, narrowSizes, maps, opType);
       encodedRhs =
           padAndSetEncoding(rewriter, loc, rhs, IREE::Encoding::MATMUL_RHS,
-                            elemTypes, narrowSizes, maps);
+                            elemTypes, narrowSizes, maps, opType);
       encodedOut =
           padAndSetEncoding(rewriter, loc, out, IREE::Encoding::MATMUL_RESULT,
-                            elemTypes, narrowSizes, maps);
+                            elemTypes, narrowSizes, maps, opType);
     } else {
       auto setEncodingWrapper = [&](Value src, int64_t operandIndex) -> Value {
         SmallVector<int64_t> roundDimsTo(linalgOp.getNumLoops(), padFactor);
         auto encoding = EncodingAttr::get(
-            linalgOp.getContext(), operandIndex, elemTypes, src.getType(),
-            narrowSizes.M, narrowSizes.N, maps, roundDimsTo);
+            linalgOp.getContext(), operandIndex, opType, elemTypes,
+            src.getType(), narrowSizes.M, narrowSizes.N, maps, roundDimsTo);
         return setEncoding(rewriter, loc, src, encoding);
       };
       encodedLhs = setEncodingWrapper(lhs, IREE::Encoding::MATMUL_LHS);
diff --git a/compiler/src/iree/compiler/GlobalOptimization/test/materialize_homogeneous_encodings.mlir b/compiler/src/iree/compiler/GlobalOptimization/test/materialize_homogeneous_encodings.mlir
index c1f4314..c0d1597 100644
--- a/compiler/src/iree/compiler/GlobalOptimization/test/materialize_homogeneous_encodings.mlir
+++ b/compiler/src/iree/compiler/GlobalOptimization/test/materialize_homogeneous_encodings.mlir
@@ -13,15 +13,15 @@
     %c1 = arith.constant 1 : index
     %dim = tensor.dim %arg0, %c0 : tensor<?x?xf32>
     %dim_0 = tensor.dim %arg0, %c1 : tensor<?x?xf32>
-    %0:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> index, index
+    %0:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> index, index
     %1 = affine.apply #map()[%0#0, %dim]
     %2 = affine.apply #map()[%0#1, %dim_0]
     %padded = tensor.pad %arg0 low[0, 0] high[%1, %2] {
     ^bb0(%arg1: index, %arg2: index):
       tensor.yield %cst : f32
     } : tensor<?x?xf32> to tensor<?x?xf32>
-    %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>>
-    %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> tensor<?x?xf32>
+    %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>>
+    %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> tensor<?x?xf32>
     util.return %4 : tensor<?x?xf32>
   }
 }
@@ -44,15 +44,15 @@
     %c1 = arith.constant 1 : index
     %dim = tensor.dim %arg0, %c0 : tensor<?x?xf32>
     %dim_0 = tensor.dim %arg0, %c1 : tensor<?x?xf32>
-    %0:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> index, index
+    %0:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> index, index
     %1 = affine.apply #map()[%0#0, %dim]
     %2 = affine.apply #map()[%0#1, %dim_0]
     %padded = tensor.pad %arg0 low[0, 0] high[%1, %2] {
     ^bb0(%arg1: index, %arg2: index):
       tensor.yield %cst : f32
     } : tensor<?x?xf32> to tensor<?x?xf32>
-    %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>>
-    %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> tensor<?x?xf32>
+    %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>>
+    %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> tensor<?x?xf32>
     util.return %4 : tensor<?x?xf32>
   }
 }
@@ -79,15 +79,15 @@
     %c1 = arith.constant 1 : index
     %dim = tensor.dim %arg0, %c0 : tensor<?x?xf32>
     %dim_0 = tensor.dim %arg0, %c1 : tensor<?x?xf32>
-    %0:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> index, index
+    %0:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> index, index
     %1 = affine.apply #map()[%0#0, %dim]
     %2 = affine.apply #map()[%0#1, %dim_0]
     %padded = tensor.pad %arg0 low[0, 0] high[%1, %2] {
     ^bb0(%arg1: index, %arg2: index):
       tensor.yield %cst : f32
     } : tensor<?x?xf32> to tensor<?x?xf32>
-    %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>>
-    %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> tensor<?x?xf32>
+    %3 = iree_encoding.set_encoding %padded : tensor<?x?xf32> -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>>
+    %4 = iree_encoding.unset_encoding %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#map1, #map2, #map3]>> -> tensor<?x?xf32>
     util.return %4 : tensor<?x?xf32>
   }
 }
diff --git a/compiler/src/iree/compiler/GlobalOptimization/test/set_encoding.mlir b/compiler/src/iree/compiler/GlobalOptimization/test/set_encoding.mlir
index 042357e..e3e96d7 100644
--- a/compiler/src/iree/compiler/GlobalOptimization/test/set_encoding.mlir
+++ b/compiler/src/iree/compiler/GlobalOptimization/test/set_encoding.mlir
@@ -18,27 +18,27 @@
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C100]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<100x250xf32> to tensor<?x?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<100x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<100x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C250]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C500]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<250x500xf32> to tensor<?x?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<250x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<250x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C100]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]]]
 //      CHECK:       tensor<100x500xf32> to tensor<?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[MATMUL:.+]] = linalg.matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -55,11 +55,11 @@
 // PAD-WITHIN-ENCODING-SAME:   %[[ARG1:[a-zA-Z0-9]+]]
 // PAD-WITHIN-ENCODING-SAME:   %[[ARG2:[a-zA-Z0-9]+]]
 // PAD-WITHIN-ENCODING:        %[[LHS:.+]] = iree_encoding.set_encoding %[[ARG0]]
-// PAD-WITHIN-ENCODING-SAME:     tensor<100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<100x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]], round_dims_to = array<i64: 16, 16, 16>>>
+// PAD-WITHIN-ENCODING-SAME:     tensor<100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<100x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]], round_dims_to = array<i64: 16, 16, 16>>>
 // PAD-WITHIN-ENCODING:        %[[RHS:.+]] = iree_encoding.set_encoding %[[ARG1]]
-// PAD-WITHIN-ENCODING-SAME:     tensor<250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<250x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]], round_dims_to = array<i64: 16, 16, 16>>>
+// PAD-WITHIN-ENCODING-SAME:     tensor<250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<250x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]], round_dims_to = array<i64: 16, 16, 16>>>
 // PAD-WITHIN-ENCODING:        %[[LHS:.+]] = iree_encoding.set_encoding %[[ARG2]]
-// PAD-WITHIN-ENCODING-SAME:     tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]], round_dims_to = array<i64: 16, 16, 16>>>
+// PAD-WITHIN-ENCODING-SAME:     tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]], round_dims_to = array<i64: 16, 16, 16>>>
 
 // -----
 
@@ -77,7 +77,7 @@
 // CHECK-SAME:     %[[ARG0:.+]]: tensor<?x?xf32>, %[[ARG1:.+]]: tensor<?x?xf32>, %[[ARG2:.+]]: tensor<?x?xf32>
 //  CHECK-DAG:     %[[C0:.+]] = arith.constant 0 : index
 //  CHECK-DAG:     %[[C1:.+]] = arith.constant 1 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[LHS_DIM0:.+]] = tensor.dim %[[ARG0]], %[[C0]]
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[LHS_DIM0]]]
 //      CHECK:   %[[LHS_DIM1:.+]] = tensor.dim %[[ARG0]], %[[C1]]
@@ -85,8 +85,8 @@
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<?x?xf32> to tensor<?x?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[RHS_DIM0:.+]] = tensor.dim %[[ARG1]], %[[C0]]
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[RHS_DIM0]]]
 //      CHECK:   %[[RHS_DIM1:.+]] = tensor.dim %[[ARG1]], %[[C1]]
@@ -94,8 +94,8 @@
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<?x?xf32> to tensor<?x?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[OUTS_DIM0:.+]] = tensor.dim %[[ARG2]], %[[C0]]
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[OUTS_DIM0]]]
 //      CHECK:   %[[OUTS_DIM1:.+]] = tensor.dim %[[ARG2]], %[[C1]]
@@ -103,7 +103,7 @@
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]]]
 //      CHECK:       tensor<?x?xf32> to tensor<?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[MATMUL:.+]] = linalg.matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -126,21 +126,21 @@
 // CHECK-SAME:     %[[ARG0:.+]]: tensor<100x250xi8>
 // CHECK-SAME:     %[[ARG1:.+]]: tensor<250x500xi8>
 // CHECK-SAME:     %[[ARG2:.+]]: tensor<100x500xi32>
-//      CHECK:   %[[LHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xi8, #iree_encoding.encoding<operand_index = 0 : index, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xi8, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high
 //      CHECK:       tensor<100x250xi8> to tensor<?x?xi8>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0 : index, element_types = [i8, i8, i32], original_type = tensor<100x250xi8>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xi8, #iree_encoding.encoding<operand_index = 1 : index, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xi8, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<100x250xi8>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xi8, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0] high
 //      CHECK:       tensor<250x500xi8> to tensor<?x?xi8>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1 : index, element_types = [i8, i8, i32], original_type = tensor<250x500xi8>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xi32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xi8, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<250x500xi8>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xi32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high
 //      CHECK:       tensor<100x500xi32> to tensor<?x?xi32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [i8, i8, i32], original_type = tensor<100x500xi32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+// CHECK-SAME:       tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<100x500xi32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
 //      CHECK:   %[[MATMUL:.+]] = linalg.matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -163,21 +163,21 @@
 // CHECK-SAME:     %[[ARG0:.+]]: tensor<100x250xf16>
 // CHECK-SAME:     %[[ARG1:.+]]: tensor<250x500xf16>
 // CHECK-SAME:     %[[ARG2:.+]]: tensor<100x500xf32>
-//      CHECK:   %[[LHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high
 //      CHECK:       tensor<100x250xf16> to tensor<?x?xf16>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f16, f16, f32], original_type = tensor<100x250xf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f16, f16, f32], original_type = tensor<100x250xf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0] high
 //      CHECK:       tensor<250x500xf16> to tensor<?x?xf16>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f16, f16, f32], original_type = tensor<250x500xf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f16, f16, f32], original_type = tensor<250x500xf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high
 //      CHECK:       tensor<100x500xf32> to tensor<?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f16, f16, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f16, f16, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
 //      CHECK:   %[[MATMUL:.+]] = linalg.matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -200,21 +200,21 @@
 // CHECK-SAME:     %[[ARG0:.+]]: tensor<100x250xf16>
 // CHECK-SAME:     %[[ARG1:.+]]: tensor<250x500xf16>
 // CHECK-SAME:     %[[ARG2:.+]]: tensor<100x500xf16>
-//      CHECK:   %[[LHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high
 //      CHECK:       tensor<100x250xf16> to tensor<?x?xf16>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f16, f16, f16], original_type = tensor<100x250xf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f16, f16, f16], original_type = tensor<100x250xf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0] high
 //      CHECK:       tensor<250x500xf16> to tensor<?x?xf16>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f16, f16, f16], original_type = tensor<250x500xf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf16, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f16, f16, f16], original_type = tensor<250x500xf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf16, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high
 //      CHECK:       tensor<100x500xf16> to tensor<?x?xf16>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f16, f16, f16], original_type = tensor<100x500xf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+// CHECK-SAME:       tensor<?x?xf16, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f16, f16, f16], original_type = tensor<100x500xf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
 //      CHECK:   %[[MATMUL:.+]] = linalg.matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -237,21 +237,21 @@
 // CHECK-SAME:     %[[ARG0:.+]]: tensor<100x250xbf16>
 // CHECK-SAME:     %[[ARG1:.+]]: tensor<250x500xbf16>
 // CHECK-SAME:     %[[ARG2:.+]]: tensor<100x500xf32>
-//      CHECK:   %[[LHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xbf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xbf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high
 //      CHECK:       tensor<100x250xbf16> to tensor<?x?xbf16>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [bf16, bf16, f32], original_type = tensor<100x250xbf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xbf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [bf16, bf16, f32], original_type = tensor<100x250xbf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xbf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0] high
 //      CHECK:       tensor<250x500xbf16> to tensor<?x?xbf16>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [bf16, bf16, f32], original_type = tensor<250x500xbf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [bf16, bf16, f32], original_type = tensor<250x500xbf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high
 //      CHECK:       tensor<100x500xf32> to tensor<?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [bf16, bf16, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [bf16, bf16, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
 //      CHECK:   %[[MATMUL:.+]] = linalg.matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -274,21 +274,21 @@
 // CHECK-SAME:     %[[ARG0:.+]]: tensor<100x250xbf16>
 // CHECK-SAME:     %[[ARG1:.+]]: tensor<250x500xbf16>
 // CHECK-SAME:     %[[ARG2:.+]]: tensor<100x500xbf16>
-//      CHECK:   %[[LHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xbf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xbf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high
 //      CHECK:       tensor<100x250xbf16> to tensor<?x?xbf16>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [bf16, bf16, bf16], original_type = tensor<100x250xbf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xbf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<100x250xbf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xbf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0] high
 //      CHECK:       tensor<250x500xbf16> to tensor<?x?xbf16>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [bf16, bf16, bf16], original_type = tensor<250x500xbf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xbf16, #iree_encoding.encoding<operand_index = 2 : index, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<250x500xbf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xbf16, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>> -> index, index
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high
 //      CHECK:       tensor<100x500xbf16> to tensor<?x?xbf16>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2 : index, element_types = [bf16, bf16, bf16], original_type = tensor<100x500xbf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+// CHECK-SAME:       tensor<?x?xbf16, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<100x500xbf16>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
 //      CHECK:   %[[MATMUL:.+]] = linalg.matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -316,30 +316,30 @@
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[LHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#2, %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]], %[[LHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x250xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<64x100x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<64x100x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[RHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]], %[[RHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x250x500xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<64x250x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<64x250x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]], %[[OUTS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x500xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<64x100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<64x100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[BATCH_MATMUL:.+]] = linalg.batch_matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -364,7 +364,7 @@
 //  CHECK-DAG:     %[[C0:.+]] = arith.constant 0 : index
 //  CHECK-DAG:     %[[C1:.+]] = arith.constant 1 : index
 //  CHECK-DAG:     %[[C2:.+]] = arith.constant 2 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[LHS_DIM0:.+]] = tensor.dim %[[ARG0]], %[[C0]]
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[LHS_DIM0]]]
 //      CHECK:   %[[LHS_DIM1:.+]] = tensor.dim %[[ARG0]], %[[C1]]
@@ -374,8 +374,8 @@
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]], %[[LHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<?x?x?xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[RHS_DIM0:.+]] = tensor.dim %[[ARG1]], %[[C0]]
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[RHS_DIM0]]]
 //      CHECK:   %[[RHS_DIM1:.+]] = tensor.dim %[[ARG1]], %[[C1]]
@@ -385,8 +385,8 @@
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]], %[[RHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<?x?x?xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[OUTS_DIM0:.+]] = tensor.dim %[[ARG2]], %[[C0]]
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[OUTS_DIM0]]]
 //      CHECK:   %[[OUTS_DIM1:.+]] = tensor.dim %[[ARG2]], %[[C1]]
@@ -396,7 +396,7 @@
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]], %[[OUTS_PADDING_SIZE2]]]
 //      CHECK:       tensor<?x?x?xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[BATCH_MATMUL:.+]] = linalg.batch_matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -424,30 +424,30 @@
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[LHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#2, %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]], %[[LHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x250xf16> to tensor<?x?x?xf16>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f16, f16, f16], original_type = tensor<64x100x250xf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f16, f16, f16], original_type = tensor<64x100x250xf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[RHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]], %[[RHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x250x500xf16> to tensor<?x?x?xf16>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f16, f16, f16], original_type = tensor<64x250x500xf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xf16, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f16, f16, f16], original_type = tensor<64x250x500xf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xf16, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f16, f16, f16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]], %[[OUTS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x500xf16> to tensor<?x?x?xf16>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf16, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f16, f16, f16], original_type = tensor<64x100x500xf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?x?xf16, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f16, f16, f16], original_type = tensor<64x100x500xf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[BATCH_MATMUL:.+]] = linalg.batch_matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -475,30 +475,30 @@
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[LHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#2, %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]], %[[LHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x250xf16> to tensor<?x?x?xf16>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f16, f16, f32], original_type = tensor<64x100x250xf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f16, f16, f32], original_type = tensor<64x100x250xf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[RHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]], %[[RHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x250x500xf16> to tensor<?x?x?xf16>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f16, f16, f32], original_type = tensor<64x250x500xf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f16, f16, f32], original_type = tensor<64x250x500xf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f16, f16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]], %[[OUTS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x500xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f16, f16, f32], original_type = tensor<64x100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f16, f16, f32], original_type = tensor<64x100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[BATCH_MATMUL:.+]] = linalg.batch_matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -526,30 +526,30 @@
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xbf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xbf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[LHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#2, %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]], %[[LHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x250xbf16> to tensor<?x?x?xbf16>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xbf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [bf16, bf16, bf16], original_type = tensor<64x100x250xbf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xbf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xbf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<64x100x250xbf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xbf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[RHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]], %[[RHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x250x500xbf16> to tensor<?x?x?xbf16>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xbf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [bf16, bf16, bf16], original_type = tensor<64x250x500xbf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xbf16, #iree_encoding.encoding<operand_index = 2 : index, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xbf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<64x250x500xbf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xbf16, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [bf16, bf16, bf16], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]], %[[OUTS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x500xbf16> to tensor<?x?x?xbf16>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xbf16, #iree_encoding.encoding<operand_index = 2 : index, element_types = [bf16, bf16, bf16], original_type = tensor<64x100x500xbf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?x?xbf16, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [bf16, bf16, bf16], original_type = tensor<64x100x500xbf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[BATCH_MATMUL:.+]] = linalg.batch_matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -577,30 +577,30 @@
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xbf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xbf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[LHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#2, %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]], %[[LHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x250xbf16> to tensor<?x?x?xbf16>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xbf16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [bf16, bf16, f32], original_type = tensor<64x100x250xbf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xbf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xbf16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [bf16, bf16, f32], original_type = tensor<64x100x250xbf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xbf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[RHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]], %[[RHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x250x500xbf16> to tensor<?x?x?xbf16>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xbf16, #iree_encoding.encoding<operand_index = 1 : index, element_types = [bf16, bf16, f32], original_type = tensor<64x250x500xbf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xbf16, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [bf16, bf16, f32], original_type = tensor<64x250x500xbf16>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [bf16, bf16, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]], %[[OUTS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x500xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [bf16, bf16, f32], original_type = tensor<64x100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [bf16, bf16, f32], original_type = tensor<64x100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[BATCH_MATMUL:.+]] = linalg.batch_matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -628,30 +628,30 @@
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xi8, #iree_encoding.encoding<operand_index = 0 : index, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x250xi8, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[LHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#2, %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]], %[[LHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x250xi8> to tensor<?x?x?xi8>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 0 : index, element_types = [i8, i8, i32], original_type = tensor<64x100x250xi8>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xi8, #iree_encoding.encoding<operand_index = 1 : index, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<64x100x250xi8>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x250x500xi8, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[RHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]], %[[RHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x250x500xi8> to tensor<?x?x?xi8>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 1 : index, element_types = [i8, i8, i32], original_type = tensor<64x250x500xi8>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xi32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?x?xi8, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<64x250x500xi8>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<64x100x500xi32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [i8, i8, i32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C64]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]], %[[OUTS_PADDING_SIZE2]]]
 //      CHECK:       tensor<64x100x500xi32> to tensor<?x?x?xi32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [i8, i8, i32], original_type = tensor<64x100x500xi32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?x?xi32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [i8, i8, i32], original_type = tensor<64x100x500xi32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[BATCH_MATMUL:.+]] = linalg.batch_matmul
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -677,25 +677,25 @@
 // CHECK-SAME:     %[[ARG2:.+]]: tensor<500xf32>
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]] = iree_encoding.upper_bound_tile_size tensor<250xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]] = iree_encoding.upper_bound_tile_size tensor<250xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index
 //      CHECK:   %[[LHS_PADDING_SIZE:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]], %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0] high[%[[LHS_PADDING_SIZE]]]
 //      CHECK:       tensor<250xf32> to tensor<?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<250xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<250xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C250]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C500]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<250x500xf32> to tensor<?x?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<250x500xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]] = iree_encoding.upper_bound_tile_size tensor<500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<250x500xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]] = iree_encoding.upper_bound_tile_size tensor<500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index
 //      CHECK:   %[[OUTS_PADDING_SIZE:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]], %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0] high[%[[OUTS_PADDING_SIZE]]]
 //      CHECK:       tensor<500xf32> to tensor<?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<500xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<500xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[VECMAT:.+]] = linalg.vecmat
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -721,25 +721,25 @@
 // CHECK-SAME:     %[[ARG2:.+]]: tensor<100xf32>
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C100]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<100x250xf32> to tensor<?x?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<100x250xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]] = iree_encoding.upper_bound_tile_size tensor<250xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<100x250xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]] = iree_encoding.upper_bound_tile_size tensor<250xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index
 //      CHECK:   %[[RHS_PADDING_SIZE:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]], %[[C250]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0] high[%[[RHS_PADDING_SIZE]]]
 //      CHECK:       tensor<250xf32> to tensor<?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<250xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]] = iree_encoding.upper_bound_tile_size tensor<100xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index
+// CHECK-SAME:       tensor<?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<250xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]] = iree_encoding.upper_bound_tile_size tensor<100xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index
 //      CHECK:   %[[OUTS_PADDING_SIZE:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]], %[[C100]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0] high[%[[OUTS_PADDING_SIZE]]]
 //      CHECK:       tensor<100xf32> to tensor<?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<100xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<100xf32>, matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[MATVEC:.+]] = linalg.matvec
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -766,28 +766,28 @@
 //  CHECK-DAG:     %[[C3:.+]] = arith.constant 3 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<3x250xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<3x250xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C3]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<3x250xf32> to tensor<?x?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<3x250xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<3x250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<3x250xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<3x250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C3]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[RHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]], %[[RHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<3x250x500xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<3x250x500xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<3x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<3x250x500xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<3x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C3]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]]]
 //      CHECK:       tensor<3x500xf32> to tensor<?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<3x500xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<3x500xf32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[VECMAT:.+]] = linalg.batch_vecmat
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -812,7 +812,7 @@
 //  CHECK-DAG:     %[[C0:.+]] = arith.constant 0 : index
 //  CHECK-DAG:     %[[C1:.+]] = arith.constant 1 : index
 //  CHECK-DAG:     %[[C2:.+]] = arith.constant 2 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[LHS_DIM0:.+]] = tensor.dim %[[ARG0]], %[[C0]]
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[LHS_DIM0]]]
 //      CHECK:   %[[LHS_DIM1:.+]] = tensor.dim %[[ARG0]], %[[C1]]
@@ -822,8 +822,8 @@
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]], %[[LHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<?x?x?xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[RHS_DIM0:.+]] = tensor.dim %[[ARG1]], %[[C0]]
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[RHS_DIM0]]]
 //      CHECK:   %[[RHS_DIM1:.+]] = tensor.dim %[[ARG1]], %[[C1]]
@@ -831,8 +831,8 @@
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<?x?xf32> to tensor<?x?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[OUTS_DIM0:.+]] = tensor.dim %[[ARG2]], %[[C0]]
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[OUTS_DIM0]]]
 //      CHECK:   %[[OUTS_DIM1:.+]] = tensor.dim %[[ARG2]], %[[C1]]
@@ -840,7 +840,7 @@
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]]]
 //      CHECK:       tensor<?x?xf32> to tensor<?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[BATCH_MATVEC:.+]] = linalg.batch_matvec
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -851,24 +851,24 @@
 // -----
 
 util.func public @fold_fill_with_set_encoding(%arg0 : index, %arg1 : index)
-  -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>> {
+  -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>> {
   %cst = arith.constant 0.0 : f32
   %0 = tensor.empty(%arg0, %arg1) : tensor<?x?xf32>
   %1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<?x?xf32>) -> tensor<?x?xf32>
   %2 = iree_encoding.set_encoding %1 : tensor<?x?xf32>
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
-  util.return %2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, element_types = [f32, f32, f32]>>
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
+  util.return %2 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 //      CHECK: util.func public @fold_fill_with_set_encoding(
-//      CHECK:   %[[EMPTY:.+]] = tensor.empty(%{{.+}}, %{{.+}}) : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>
+//      CHECK:   %[[EMPTY:.+]] = tensor.empty(%{{.+}}, %{{.+}}) : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 //      CHECK:   %[[FILL:.+]] = linalg.fill
-// CHECK-SAME:       outs(%[[EMPTY]] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, element_types = [f32, f32, f32]>>)
+// CHECK-SAME:       outs(%[[EMPTY]] : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>>)
 //      CHECK:   util.return %[[FILL]]
 
 // -----
 
 util.func public @fold_fill_with_tensor_pad(%arg0 : index, %arg1 : index, %arg2 : index, %arg3 : index)
-    -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>> {
+    -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>> {
   %cst = arith.constant 0.0 : f32
   %0 = tensor.empty(%arg0, %arg1) : tensor<?x?xf32>
   %1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<?x?xf32>) -> tensor<?x?xf32>
@@ -877,12 +877,12 @@
     tensor.yield %cst : f32
   } : tensor<?x?xf32> to tensor<?x?xf32>
   %3 = iree_encoding.set_encoding %2 : tensor<?x?xf32>
-      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>
-  util.return %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, element_types = [f32, f32, f32]>>
+      -> tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>
+  util.return %3 : tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2, op_type = matmul, element_types = [f32, f32, f32]>>
 }
 //      CHECK: util.func public @fold_fill_with_tensor_pad(
 //      CHECK:   %[[EMPTY:.+]] = tensor.empty(
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : i64, element_types = [f32, f32, f32]>>
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : i64, op_type = matmul, element_types = [f32, f32, f32]>>
 //      CHECK:   %[[FILL:.+]] = linalg.fill
 // CHECK-SAME:       outs(%[[EMPTY]] :
 //      CHECK:   util.return %[[FILL]]
@@ -960,9 +960,9 @@
 //  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1, d2) -> (d2, d1)>
 //  CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0, d1, d2) -> (d0, d1)>
 //      CHECK: util.func public @matmul_casted_from_i1_f32f32f32
-// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<64x256xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<256x128xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<64x128xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<64x256xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<256x128xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<64x128xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
 
 // -----
 
@@ -988,9 +988,9 @@
 //  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1, d2) -> (d2, d1)>
 //  CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0, d1, d2) -> (d0, d1)>
 //      CHECK: util.func public @matmul_generic_casted_from_i1_f32f32f32
-// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<64x256xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<256x128xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<64x128xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<64x256xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<256x128xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+// CHECK:         set_encoding {{.+}} tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<64x128xf32>, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
 
 // -----
 
@@ -1004,9 +1004,9 @@
 //  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1, d2) -> (d2, d1)>
 //  CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0, d1, d2) -> (d0, d1)>
 //      CHECK: util.func public @matmul_f32f32f32_narrow_M(
-//      CHECK:  iree_encoding.upper_bound_tile_size tensor<2x250xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], matmul_narrow_M = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:  iree_encoding.upper_bound_tile_size tensor<250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], matmul_narrow_M = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:  iree_encoding.upper_bound_tile_size tensor<2x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], matmul_narrow_M = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:  iree_encoding.upper_bound_tile_size tensor<2x250xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_M = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:  iree_encoding.upper_bound_tile_size tensor<250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_M = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:  iree_encoding.upper_bound_tile_size tensor<2x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_M = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
 //      CHECK:   linalg.matmul
 
 // -----
@@ -1021,9 +1021,9 @@
 //  CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1, d2, d3) -> (d0, d3, d2)>
 //  CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2)>
 //      CHECK: util.func public @batch_matmul_f32f32f32_narrow_MN(
-//      CHECK:   iree_encoding.upper_bound_tile_size tensor<64x4x250xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], matmul_narrow_N = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   iree_encoding.upper_bound_tile_size tensor<64x250x2xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], matmul_narrow_N = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
-//      CHECK:   iree_encoding.upper_bound_tile_size tensor<64x4x2xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], matmul_narrow_N = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   iree_encoding.upper_bound_tile_size tensor<64x4x250xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   iree_encoding.upper_bound_tile_size tensor<64x250x2xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
+//      CHECK:   iree_encoding.upper_bound_tile_size tensor<64x4x2xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], matmul_narrow_N = 2 : index, user_indexing_maps = [#[[MAP0]], #[[MAP1]], #[[MAP2]]]>>
 //      CHECK:   linalg.batch_matmul
 
 // -----
@@ -1046,27 +1046,27 @@
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<250x100xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<250x100xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C250]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<250x100xf32> to tensor<?x?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<250x100xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<250x100xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C250]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C500]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<250x500xf32> to tensor<?x?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<250x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<250x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C100]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]]]
 //      CHECK:       tensor<100x500xf32> to tensor<?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[MATMUL:.+]] = linalg.matmul_transpose_a
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -1093,27 +1093,27 @@
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C100]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<100x250xf32> to tensor<?x?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<100x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<500x250xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<100x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<500x250xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C500]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<500x250xf32> to tensor<?x?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<500x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<500x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C100]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]]]
 //      CHECK:       tensor<100x500xf32> to tensor<?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[MATMUL:.+]] = linalg.matmul_transpose_b
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -1141,30 +1141,30 @@
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x250x100xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x250x100xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C2]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[LHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#2, %[[C100]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]], %[[LHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<2x250x100xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<2x250x100xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x250x100xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x250x500xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C2]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C250]]]
 //      CHECK:   %[[RHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]], %[[RHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<2x250x500xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<2x250x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x250x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C2]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]], %[[OUTS_PADDING_SIZE2]]]
 //      CHECK:       tensor<2x100x500xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<2x100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[BATCH_MATMUL:.+]] = linalg.batch_matmul_transpose_a
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -1192,30 +1192,30 @@
 //  CHECK-DAG:     %[[C100:.+]] = arith.constant 100 : index
 //  CHECK-DAG:     %[[C250:.+]] = arith.constant 250 : index
 //  CHECK-DAG:     %[[C500:.+]] = arith.constant 500 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x100x250xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C2]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[LHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#2, %[[C250]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]], %[[LHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<2x100x250xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, element_types = [f32, f32, f32], original_type = tensor<2x100x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x500x250xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x100x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x500x250xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C2]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C500]]]
 //      CHECK:   %[[RHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#2, %[[C250]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]], %[[RHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<2x500x250xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, element_types = [f32, f32, f32], original_type = tensor<2x500x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x500x250xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<2x100x500xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C2]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C100]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#2, %[[C500]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]], %[[OUTS_PADDING_SIZE2]]]
 //      CHECK:       tensor<2x100x500xf32> to tensor<?x?x?xf32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [f32, f32, f32], original_type = tensor<2x100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?x?xf32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32], original_type = tensor<2x100x500xf32>, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[BATCH_MATMUL:.+]] = linalg.batch_matmul_transpose_b
 // CHECK-SAME:       ins(%[[LHS]], %[[RHS]] :
 // CHECK-SAME:       outs(%[[OUTS]] :
@@ -1249,28 +1249,28 @@
 //  CHECK-DAG:     %[[C4096:.+]] = arith.constant 4096 : index
 //  CHECK-DAG:     %[[C128:.+]] = arith.constant 128 : index
 //  CHECK-DAG:     %[[C32:.+]] = arith.constant 32 : index
-//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<32x128xi16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+//      CHECK:   %[[LHS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<32x128xi16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[LHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#0, %[[C32]]]
 //      CHECK:   %[[LHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[LHS_TILE_SIZE]]#1, %[[C128]]]
 //      CHECK:   %[[LHS_PAD:.+]] = tensor.pad %[[ARG0]] low[0, 0] high[%[[LHS_PADDING_SIZE0]], %[[LHS_PADDING_SIZE1]]]
 //      CHECK:       tensor<32x128xi16> to tensor<?x?xi16>
 //      CHECK:   %[[LHS:.+]] = iree_encoding.set_encoding %[[LHS_PAD]]
-// CHECK-SAME:       tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0 : index, element_types = [i16, ui4, i32], original_type = tensor<32x128xi16>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<4096x32x128xi4, #iree_encoding.encoding<operand_index = 1 : index, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
+// CHECK-SAME:       tensor<?x?xi16, #iree_encoding.encoding<operand_index = 0 : index, op_type = matmul, element_types = [i16, ui4, i32], original_type = tensor<32x128xi16>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[RHS_TILE_SIZE:.+]]:3 = iree_encoding.upper_bound_tile_size tensor<4096x32x128xi4, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index, index
 //      CHECK:   %[[RHS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#0, %[[C4096]]]
 //      CHECK:   %[[RHS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#1, %[[C32]]]
 //      CHECK:   %[[RHS_PADDING_SIZE2:.+]] = affine.apply #[[MAP]]()[%[[RHS_TILE_SIZE]]#2, %[[C128]]]
 //      CHECK:   %[[RHS_PAD:.+]] = tensor.pad %[[ARG1]] low[0, 0, 0] high[%[[RHS_PADDING_SIZE0]], %[[RHS_PADDING_SIZE1]], %[[RHS_PADDING_SIZE2]]]
 //      CHECK:       tensor<4096x32x128xi4> to tensor<?x?x?xi4>
 //      CHECK:   %[[RHS:.+]] = iree_encoding.set_encoding %[[RHS_PAD]]
-// CHECK-SAME:       tensor<?x?x?xi4, #iree_encoding.encoding<operand_index = 1 : index, element_types = [i16, ui4, i32], original_type = tensor<4096x32x128xi4>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
-//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<4096x32xi32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
+// CHECK-SAME:       tensor<?x?x?xi4, #iree_encoding.encoding<operand_index = 1 : index, op_type = matmul, element_types = [i16, ui4, i32], original_type = tensor<4096x32x128xi4>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+//      CHECK:   %[[OUTS_TILE_SIZE:.+]]:2 = iree_encoding.upper_bound_tile_size tensor<4096x32xi32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [i16, ui4, i32], matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>> -> index, index
 //      CHECK:   %[[OUTS_PADDING_SIZE0:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#0, %[[C4096]]]
 //      CHECK:   %[[OUTS_PADDING_SIZE1:.+]] = affine.apply #[[MAP]]()[%[[OUTS_TILE_SIZE]]#1, %[[C32]]]
 //      CHECK:   %[[OUTS_PAD:.+]] = tensor.pad %[[ARG2]] low[0, 0] high[%[[OUTS_PADDING_SIZE0]], %[[OUTS_PADDING_SIZE1]]]
 //      CHECK:       tensor<4096x32xi32> to tensor<?x?xi32>
 //      CHECK:   %[[OUTS:.+]] = iree_encoding.set_encoding %[[OUTS_PAD]]
-// CHECK-SAME:       tensor<?x?xi32> -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2 : index, element_types = [i16, ui4, i32], original_type = tensor<4096x32xi32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
+// CHECK-SAME:       tensor<?x?xi32> -> tensor<?x?xi32, #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [i16, ui4, i32], original_type = tensor<4096x32xi32>, matmul_narrow_M = 1 : index, user_indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]>>
 //      CHECK:   %[[GENERIC:.+]] = linalg.generic
 // CHECK-SAME:       indexing_maps = [#[[MAP1]], #[[MAP2]], #[[MAP3]]]
 // CHECK-SAME:       iterator_types = ["parallel", "parallel", "reduction"]