[Mlir-commits] [mlir] [MLIR][Vector] Add DropUnitDimFromBroadcastOp pattern (PR #92938)

[Benjamin Maxwell]

================
@@ -1703,6 +1703,66 @@ struct DropUnitDimFromElementwiseOps final
   }
 };
 
+/// Drops unit non scalable dimensions inside a broadcastOp which are shared
+/// among source and result with shape_casts.
+/// The newly inserted shape_cast Ops fold (before Op) and then
+/// restore the unit dim after Op. Source type is required to be a vector.
+///
+/// Ex:
+/// ```
+///  %bc = vector.broadcast %arg0 : vector<1x4xf32> to vector<1x3x1x4xf32>
+///  %cast = vector.shape_cast %bc : vector<1x3x1x4xf32> to vector<1x3x4xf32>
+/// ```
+///
+/// Gets converted to:
+///
+/// ```
+///  %sc_arg = vector.shape_cast %arg0 : vector<1x4xf32> to vector<4xf32>
+///  %bc = vector.broadcast %arg : vector<4xf32> to vector<1x3x4xf32>
+///  %cast_new = vector.shape_cast %bc : vector<1x3x4xf32> to
+///    vector<1x3x1x4xf32>
+///  %cast = vector.shape_cast %cast_new : vector<1x3x1x4xf32> to
+///    vector<1x3x4xf32>
+/// ```
+/// %cast_new and %cast can be folded away.
+struct DropUnitDimFromBroadcastOp final
+    : public OpRewritePattern<vector::BroadcastOp> {
+  using OpRewritePattern::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(vector::BroadcastOp broadcastOp,
+                                PatternRewriter &rewriter) const override {
+    auto srcVecTy = dyn_cast<VectorType>(broadcastOp.getSourceType());
+    if (!srcVecTy)
+      return failure();
+    auto resVecTy = broadcastOp.getResultVectorType();
+    auto srcVecTyBuilder = VectorType::Builder(srcVecTy);
+    auto resVecTyBuilder = VectorType::Builder(resVecTy);
----------------
MacDue wrote:

Those `static_casts` look very suspect :sweat_smile:. It looks like that's just going to make a SmallVector of two (likely true) bools. I think keeping it as-is is fine as it's simpler, and likely not really a performance concern (vector types are normally small). 

https://github.com/llvm/llvm-project/pull/92938