[Mlir-commits] [mlir] [mlir][Vector] Add fold transpose(shape_cast) -> shape_cast (PR #73951)

[Benjamin Maxwell]

================
@@ -5548,12 +5548,55 @@ class FoldTransposeCreateMask final : public OpRewritePattern<TransposeOp> {
   }
 };
 
+/// Folds transpose(shape_cast) into a new shape_cast, when the transpose just
+/// permutes a unit dim from the result of the shape_cast.
+class FoldTransposeShapeCast : public OpRewritePattern<TransposeOp> {
+  using OpRewritePattern::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(TransposeOp transpOp,
+                                PatternRewriter &rewriter) const override {
+    Value transposeSrc = transpOp.getVector();
+    auto shapeCastOp = transposeSrc.getDefiningOp<vector::ShapeCastOp>();
+    if (!shapeCastOp)
+      return failure();
+
+    auto sourceType = transpOp.getSourceVectorType();
----------------
MacDue wrote:

The point for us is `vector<[4]x1xf32>` is an impossible type, there's no legal lowering for that in LLVM (only trailing scalable dimensions are supported). So we need a mechanism (such as this), which allows it to be eliminated. 

https://github.com/llvm/llvm-project/pull/73951