[Mlir-commits] [mlir] [MLIR][Vector] Added ToElementsOp::fold for broadcast->to_elements pattern rewrite. (PR #160318)

[Kunwar Grover]

================
@@ -2410,6 +2443,91 @@ ToElementsOp::inferReturnTypes(MLIRContext *ctx, std::optional<Location> loc,
   return success();
 }
 
+/// Canonicalize `vector.to_elements(vector.broadcast(%v))` where `%v` is a
+/// vector:
+/// - Build `vector.to_elements %v` and remap each destination element to the
+///   corresponding source element using broadcast rules (match or 1 →
+///   replicate).
+///
+/// Example:
+///   %v = vector.broadcast %src : vector<2xf32> to vector<3x2xf32>
+///   %e:6 = vector.to_elements %v : vector<3x2xf32>
+/// becomes:
+///   %src_elems:2 = vector.to_elements %src : vector<2xf32>
+///   // uses: %src_elems#0, %src_elems#1, %src_elems#0,
+///   //       %src_elems#1, %src_elems#0, %src_elems#1
+
+class ToElementsOfBroadcast final : public OpRewritePattern<ToElementsOp> {
+  using OpRewritePattern::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(ToElementsOp toElementsOp,
+                                PatternRewriter &rewriter) const override {
+    auto bcastOp = toElementsOp.getSource().getDefiningOp<BroadcastOp>();
+    if (!bcastOp)
+      return failure();
+
+    // Only handle broadcasts from a vector source here.
+    auto srcType = dyn_cast<VectorType>(bcastOp.getSource().getType());
+    if (!srcType)
+      return failure();
+
+    auto dstType = cast<VectorType>(toElementsOp.getSource().getType());
+
+    // Bail on scalable vectors.
----------------
Groverkss wrote:

Does toElementsOp actually allow scalable vectors? Remove this check if not.

https://github.com/llvm/llvm-project/pull/160318