[Mlir-commits] [mlir] [MLIR][XeVM] Update XeVM type converter (PR #189306)

[Sang Ik Lee]

================
@@ -1084,14 +1049,25 @@ struct ConvertXeGPUToXeVMPass
   using Base::Base;
 
   void runOnOperation() override {
-    LLVMTypeConverter typeConverter(&getContext());
+    MLIRContext *context = &getContext();
+
+    // XeVM type converter is based on LLVM type converter with the
+    // following customizations.
+    // First, type conversion rules are added for xegpu custom types,
+    // TensorDescType and MemDescType.
+    // Second, MemRefType is lowered to single integer type
+    // Third, VectorType of single element or 0D is converted to vector
+    // element type. Otherwise, vector type is flatten to 1D.
+    LowerToLLVMOptions options(context);
+    options.overrideIndexBitwidth(this->use64bitIndex ? 64 : 32);
+    LLVMTypeConverter typeConverter(context, options);
+
+    Type xevmIndexType = typeConverter.convertType(IndexType::get(context));
+    Type i32Type = IntegerType::get(context, 32);
     typeConverter.addConversion([&](VectorType type) -> Type {
-      unsigned rank = type.getRank();
-      auto elemType = type.getElementType();
-      // If the element type is index, convert it to i64.
-      if (llvm::isa<IndexType>(elemType))
-        elemType = IntegerType::get(&getContext(), 64);
+      auto elemType = typeConverter.convertType(type.getElementType());
----------------
silee2 wrote:

Thanks. Missed updating materialization cast.

https://github.com/llvm/llvm-project/pull/189306