Reapply "[mlir][python] allow DenseIntElementsAttr for index type (#118947)" (#124804)

This reapplies #118947 and adapts to nanobind.

Author: mgehre-amd

mlir/include/mlir-c/BuiltinAttributes.h

@@ -556,6 +556,8 @@ MLIR_CAPI_EXPORTED int64_t
 mlirDenseElementsAttrGetInt64Value(MlirAttribute attr, intptr_t pos);
 MLIR_CAPI_EXPORTED uint64_t
 mlirDenseElementsAttrGetUInt64Value(MlirAttribute attr, intptr_t pos);
+MLIR_CAPI_EXPORTED uint64_t
+mlirDenseElementsAttrGetIndexValue(MlirAttribute attr, intptr_t pos);
 MLIR_CAPI_EXPORTED float mlirDenseElementsAttrGetFloatValue(MlirAttribute attr,
                                                             intptr_t pos);
 MLIR_CAPI_EXPORTED double

mlir/lib/Bindings/Python/IRAttributes.cpp

@@ -1372,13 +1372,19 @@ class PyDenseIntElementsAttribute
 
     MlirType type = mlirAttributeGetType(*this);
     type = mlirShapedTypeGetElementType(type);
-    assert(mlirTypeIsAInteger(type) &&
-           "expected integer element type in dense int elements attribute");
+    // Index type can also appear as a DenseIntElementsAttr and therefore can be
+    // casted to integer.
+    assert(mlirTypeIsAInteger(type) ||
+           mlirTypeIsAIndex(type) && "expected integer/index element type in "
+                                     "dense int elements attribute");
     // Dispatch element extraction to an appropriate C function based on the
     // elemental type of the attribute. nb::int_ is implicitly constructible
     // from any C++ integral type and handles bitwidth correctly.
     // TODO: consider caching the type properties in the constructor to avoid
     // querying them on each element access.
+    if (mlirTypeIsAIndex(type)) {
+      return nb::int_(mlirDenseElementsAttrGetIndexValue(*this, pos));
+    }
     unsigned width = mlirIntegerTypeGetWidth(type);
     bool isUnsigned = mlirIntegerTypeIsUnsigned(type);
     if (isUnsigned) {

mlir/lib/CAPI/IR/BuiltinAttributes.cpp

@@ -758,6 +758,9 @@ int64_t mlirDenseElementsAttrGetInt64Value(MlirAttribute attr, intptr_t pos) {
 uint64_t mlirDenseElementsAttrGetUInt64Value(MlirAttribute attr, intptr_t pos) {
   return llvm::cast<DenseElementsAttr>(unwrap(attr)).getValues<uint64_t>()[pos];
 }
+uint64_t mlirDenseElementsAttrGetIndexValue(MlirAttribute attr, intptr_t pos) {
+  return llvm::cast<DenseElementsAttr>(unwrap(attr)).getValues<uint64_t>()[pos];
+}
 float mlirDenseElementsAttrGetFloatValue(MlirAttribute attr, intptr_t pos) {
   return llvm::cast<DenseElementsAttr>(unwrap(attr)).getValues<float>()[pos];
 }

mlir/test/python/dialects/builtin.py

@@ -246,3 +246,7 @@ def testDenseElementsAttr():
         # CHECK{LITERAL}: dense<[[0, 1], [2, 3]]> : tensor<2x2xi32>
         print(DenseElementsAttr.get(values, type=VectorType.get((2, 2), i32)))
         # CHECK{LITERAL}: dense<[[0, 1], [2, 3]]> : vector<2x2xi32>
+        idx_values = np.arange(4, dtype=np.int64)
+        idx_type = IndexType.get()
+        print(DenseElementsAttr.get(idx_values, type=VectorType.get([4], idx_type)))
+        # CHECK{LITERAL}: dense<[0, 1, 2, 3]> : vector<4xindex>

mlir/test/python/ir/array_attributes.py

@@ -572,6 +572,10 @@ def testGetDenseElementsIndex():
         print(arr)
         # CHECK: True
         print(arr.dtype == np.int64)
+        array = np.array([1, 2, 3], dtype=np.int64)
+        attr = DenseIntElementsAttr.get(array, type=VectorType.get([3], idx_type))
+        # CHECK: [1, 2, 3]
+        print(list(DenseIntElementsAttr(attr)))
 
 
 # CHECK-LABEL: TEST: testGetDenseResourceElementsAttr

mlir/test/python/ir/attributes.py

@@ -366,6 +366,10 @@ def testDenseIntAttr():
         # CHECK: i1
         print(ShapedType(a.type).element_type)
 
+        shape = Attribute.parse("dense<[0, 1, 2, 3]> : vector<4xindex>")
+        # CHECK: attr: dense<[0, 1, 2, 3]>
+        print("attr:", shape)
+
 
 @run
 def testDenseArrayGetItem():