[Mlir-commits] [mlir] [MLIR][LLVM][DLTI] Handle data layout token 'n32:64' (PR #141299)

[Tobias Gysi]

================
@@ -251,6 +251,30 @@ LogicalResult DataLayoutImporter::tryToEmplaceFunctionPointerAlignmentEntry(
   return success();
 }
 
+LogicalResult
+DataLayoutImporter::tryToEmplaceLegalIntWidthsEntry(StringRef token) {
+  auto key =
+      StringAttr::get(context, DLTIDialect::kDataLayoutLegalIntWidthsKey);
+  if (keyEntries.count(key))
+    return success();
+
+  FailureOr<SmallVector<uint64_t>> intWidths = tryToParseIntList(token);
+  if (failed(intWidths) || intWidths->empty())
+    return failure();
+
+  OpBuilder builder(context);
+  SmallVector<Attribute> intWidthAttrs;
+  intWidthAttrs.reserve(intWidths->size());
+  llvm::for_each(*intWidths, [&](uint64_t width) {
+    intWidthAttrs.push_back(builder.getI32IntegerAttr(width));
+  });
+
+  keyEntries.try_emplace(
+      key, DataLayoutEntryAttr::get(
+               key, ArrayAttr::get(key.getContext(), intWidthAttrs)));
----------------
gysit wrote:

Using DenseI32ArrayAttr to store integer bit widths would be nice. That way we do not need to create attributes for the individual scalar values. Note that the builder can also directly create a dense I32 array using `builder.getDenseI32ArrayAttr({1, 2, 3}))`.

https://github.com/llvm/llvm-project/pull/141299