[Mlir-commits] [mlir] [WIP][MLIR][Affine] Add vector support to affine.linearize_index and affine.delinearize_index (PR #188369)

[Keshav Vinayak Jha]

================
@@ -29,19 +31,85 @@ using namespace mlir;
 using namespace mlir::affine;
 
 namespace {
+
+/// Create a constant splat of the given type with the given integer value.
+static Value createTypedConstant(OpBuilder &b, Location loc, Type type,
+                                 int64_t value) {
+  if (auto vecTy = dyn_cast<VectorType>(type))
+    return arith::ConstantOp::create(
+        b, loc, DenseElementsAttr::get(vecTy, b.getIndexAttr(value)));
+  return arith::ConstantIndexOp::create(b, loc, value);
+}
+
+/// Materialize an OpFoldResult (which represents a scalar index or constant)
+/// as a Value matching the given target type. For vector target types, scalar
+/// constants are splatted. Returns failure for dynamic basis with vector types
+/// since that requires vector.broadcast which is not available here.
+static FailureOr<Value> materializeBasis(OpBuilder &b, Location loc,
+                                         OpFoldResult ofr, Type targetType) {
+  std::optional<int64_t> cst = getConstantIntValue(ofr);
+  if (cst)
+    return createTypedConstant(b, loc, targetType, *cst);
+  // Dynamic scalar basis value. For scalar target types, return as-is.
+  if (isa<IndexType>(targetType))
+    return getValueOrCreateConstantIndexOp(b, loc, ofr);
+  // Dynamic scalar basis with vector target type -- would need
----------------
keshavvinayak01 wrote:

Adding a dependance with vector dialect to support broadcast or `tensor.splat` can handle this case. Not sure if we should do that though.

https://github.com/llvm/llvm-project/pull/188369