[Mlir-commits] [mlir] [mlir][linalg] Restrict linalg.pack to not have artificial padding. (PR #149624)

[Han-Chung Wang]

================
@@ -220,6 +243,29 @@ struct FoldPadWithPackOp : public OpRewritePattern<PackOp> {
       if (!isEqualConstantIntOrValue(paddingValue, constantPaddingValue))
         return failure();
 
+    // Folding is not allowed if it introduces artificial padding.
+    RankedTensorType unpackedType = packOp.getSourceType();
+    SmallVector<int64_t> outerShapeWithoutTranspose =
+        getPackedOuterShapeWithoutTransposition(packOp);
+    for (auto [pos, tileSize, high] :
+         llvm::zip_equal(packOp.getInnerDimsPos(), packOp.getStaticInnerTiles(),
+                         padOp.getMixedHighPad())) {
+      if (unpackedType.isDynamicDim(pos))
+        return failure();
+      if (ShapedType::isDynamic(outerShapeWithoutTranspose[pos]))
+        return failure();
+      if (ShapedType::isDynamic(tileSize))
----------------
hanhanW wrote:

I think it is more like a dynamic check, but not a static check. I.e., it can only be asserted during runtime for most cases, IMO. E.g., you don't do any out-of-bound checks for tensor.extract_slice for dynamic sizes/dims, you do out-of-bound checks when the op has static sizes and offsets.

So I think it is user's responsibility to provide correct shape when they construct the IR.

https://github.com/llvm/llvm-project/pull/149624