[Mlir-commits] [mlir] [mlir][tensor] Extend the logic to generalise tensor.pack (PR #109815)

[Han-Chung Wang]

================
@@ -1178,10 +1212,20 @@ LogicalResult GeneralizeOuterUnitDimsPackOpPattern::matchAndRewrite(
   LLVM_DEBUG(DBGS() << "Pack permutation: " << packOp << "\n";
              llvm::interleaveComma(perm, DBGS() << "perm: "); DBGSNL(););
 
-  SmallVector<int64_t> transpShape = readShape;
-  applyPermutationToVector<int64_t>(transpShape, perm);
+  applyPermutationToVector<OpFoldResult>(transShapeForEmpty, perm);
 
-  Value empty = rewriter.create<tensor::EmptyOp>(loc, transpShape, elemType);
+  // If there's a tile with a dynamic size, retrieve its size. ATM only 1
+  // dynamic tile is allowed.
+  Value dynDimSize;
+  for (auto tile : packOp.getMixedTiles()) {
+    if (tile.is<Value>()) {
+      assert(!dynDimSize && "Only one scalable size is supported ATM.");
+      dynDimSize = cast<Value>(tile);
+    }
+  }
----------------
hanhanW wrote:

Nit: remove the dead code. We don't need this, right? There are no uses, and the function that should account this is `getPackOpSourceOrPaddedSource`.

https://github.com/llvm/llvm-project/pull/109815