[Mlir-commits] [mlir] [mlir][linalg] Use ub.poison in data layout propagation if a packed operand requires padding. (PR #159467)

[Han-Chung Wang]

================
@@ -287,12 +288,24 @@ getOrCreatePackedViewOfOperand(OpBuilder &b, Location loc, PackInfo packInfo,
   // The operand does not have dimensions that relates to pack op.
   if (innerDimsPos.empty() && outerDimsPerm.empty())
     return std::make_tuple(opOperand->get(), indexingMap);
-
+  auto inputType = cast<RankedTensorType>(opOperand->get().getType());
+  auto maybeIntInnerTileSizes = getConstantIntValues(innerTileSizes);
+  if (!maybeIntInnerTileSizes.has_value()) {
+    return failure();
+  }
+  if (!poisonPaddingOk &&
+      linalg::PackOp::requirePaddingValueStrict(
+          inputType.getShape(), innerDimsPos,
+          linalg::PackOp::inferPackedType(inputType, *maybeIntInnerTileSizes,
+                                          innerDimsPos, outerDimsPerm)
+              .getShape(),
+          outerDimsPerm, innerTileSizes))
+    return failure();
----------------
hanhanW wrote:

I think it is better to move the check outside `getOrCreatePackedViewOfOperand` function. Reason:

- We should not create operations if the pattern returns failure at the end.
- In this context, the order matters. You could create a pack op for the first operand, then get a failure from packing the second operand. We should not generate the pack op for the first operand in this case.


https://github.com/llvm/llvm-project/pull/159467