[Mlir-commits] [mlir] [mlir][vector] Add `SwapShapeCastOfTranspose` canonicalizer pattern (PR #100933)

[Benjamin Maxwell]

================
@@ -5480,12 +5480,100 @@ class ShapeCastBroadcastFolder final : public OpRewritePattern<ShapeCastOp> {
   }
 };
 
+/// Returns an iterator over the dims (inc scalability) of a VectorType.
+static auto getDims(VectorType vType) {
+  return llvm::zip_equal(vType.getShape(), vType.getScalableDims());
+}
+
+/// Helper to drop (fixed-size) unit dims from a VectorType.
+static VectorType dropUnitDims(VectorType vType) {
+  SmallVector<bool> scalableFlags;
+  SmallVector<int64_t> dimSizes;
+  for (auto dim : getDims(vType)) {
+    if (dim == std::make_tuple(1, false))
+      continue;
+    auto [size, scalableFlag] = dim;
----------------
MacDue wrote:

This is done deliberately as a more convenience way to compare dim sizes (as there's no good vector API for it). 

I prefer the tuple to `size == 1 && !scalableFlag` (which I find takes longer to grok). 

https://github.com/llvm/llvm-project/pull/100933