[mlir][vector] Disable BreakDownVectorBitCast for scalable vectors (#122725)

`BreakDownVectorBitCast` leverages
  * `vector.extract_strided_slices` + `vector.insert_strided_slices`

As these Ops do not support extracting scalable sub-vectors (i.e.
extracting/inserting a fraction of a scalable dim), it's best to bail
out.

Author: banach-space

mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp

@@ -906,6 +906,13 @@ struct BreakDownVectorBitCast : public OpRewritePattern<vector::BitCastOp> {
     VectorType castDstType = bitcastOp.getResultVectorType();
     assert(castSrcType.getRank() == castDstType.getRank());
 
+    // This transformation builds on top of
+    // vector.{extract|insert}_strided_slice, which do not support
+    // extracting/inserting "scallable sub-vectors". Bail out.
+    if (castSrcType.isScalable())
+      return rewriter.notifyMatchFailure(bitcastOp,
+                                         "Scalable vectors are not supported");
+
     // Only support rank 1 case for now.
     if (castSrcType.getRank() != 1)
       return failure();

mlir/test/Dialect/Vector/vector-break-down-bitcast.mlir

@@ -39,3 +39,14 @@ func.func @bitcast_i8_to_i32(%input: vector<16xi8>) -> vector<4xi32> {
 // CHECK: %[[CAST3:.+]] = vector.bitcast %[[EXTRACT3]] : vector<4xi8> to vector<1xi32>
 // CHECK: %[[INSERT3:.+]] = vector.insert_strided_slice %[[CAST3]], %[[INSERT2]] {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
 // CHECK: return %[[INSERT3]]
+
+// -----
+
+// Scalable vectors are not supported!
+
+// CHECK-LABEL: func.func @bitcast_scalable_negative
+// CHECK: vector.bitcast
+func.func @bitcast_scalable_negative(%input: vector<[8]xf16>) -> vector<[4]xf32> {
+  %0 = vector.bitcast %input : vector<[8]xf16> to vector<[4]xf32>
+  return %0: vector<[4]xf32>
+}