[Mlir-commits] [mlir] [mlir][vector] Add vector.step operation (PR #96776)

[Diego Caballero]

================
@@ -6316,6 +6316,20 @@ OpFoldResult SplatOp::fold(FoldAdaptor adaptor) {
   return SplatElementsAttr::get(getType(), {constOperand});
 }
 
+//===----------------------------------------------------------------------===//
+// StepOp
+//===----------------------------------------------------------------------===//
+
+OpFoldResult StepOp::fold(FoldAdaptor adaptor) {
+  auto resultType = cast<VectorType>(getType());
+  if (resultType.isScalable())
+    return nullptr;
+  SmallVector<APInt> indices;
+  for (unsigned i = 0; i < resultType.getNumElements(); i++)
+    indices.push_back(APInt(/*width=*/64, i));
+  return DenseElementsAttr::get(resultType, indices);
----------------
dcaballe wrote:

Thanks for clarifying! I think consistency is really important to make sure fixed and scalable worlds are aligned when possible and we don't end up with a different set of patterns and problems when we shouldn't. `vector.step` looks like a good abstraction for both fixed and scalable vectors so I'm +1 to removing this canonicalization pattern and let `vector.step` be the canonical representation also for fixed vectors.

https://github.com/llvm/llvm-project/pull/96776