[mlir][Vector] Fix an assertion on failing cast in vector-transfer-flatten-patterns

[bviyer]

When the result is not a vectorType, there is an assert. This patch will do the check
and bail when the result is not a VectorType.

[llvmbot]

@llvm/pr-subscribers-mlir-vector

@llvm/pr-subscribers-mlir

Author: Balaji V. Iyer. (bviyer)

Changes

When the result is not a vectorType, there is an assert. This patch will do the check
and bail when the result is not a VectorType.

---

Full diff: https://github.com/llvm/llvm-project/pull/86030.diff

1 Files Affected:

• (modified) mlir/lib/Dialect/Vector/Transforms/VectorLinearize.cpp (+2-2)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorLinearize.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorLinearize.cpp
index 7ca03537049812..38536de43f13f2 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorLinearize.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorLinearize.cpp
@@ -22,9 +22,9 @@ using namespace mlir;
 static bool isLessThanTargetBitWidth(Operation *op, unsigned targetBitWidth) {
   auto resultTypes = op->getResultTypes();
   for (auto resType : resultTypes) {
-    VectorType vecType = cast<VectorType>(resType);
+    VectorType vecType = dyn_cast<VectorType>(resType);
     // Reject index since getElementTypeBitWidth will abort for Index types.
-    if (vecType.getElementType().isIndex())
+    if (!vecType || vecType.getElementType().isIndex())
       return false;
     unsigned trailingVecDimBitWidth =
         vecType.getShape().back() * vecType.getElementTypeBitWidth();

[joker-eph]

Thanks for the fix!

Can you add a test that shows how the assert is triggered?

[bviyer]

@joker-eph I added a test that will fail without the fix.

[banach-space]

@joker-eph I added a test that will fail without the fix.

I don't hit the assert when LLVM_ENABLE_ASSERTIONS are ON. How do you trigger it?

[joker-eph]

I'm confused, the test does not match the code change: did you try running the test at HEAD and confirm it is crashing?

[github-actions]

✅ With the latest revision this PR passed the Python code formatter.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[bviyer]

I'm confused, the test does not match the code change: did you try running the test at HEAD and confirm it is crashing?

@joker-eph and @banach-space The test was correct one but I accidentally put in the wrong file. Very sorry about the mess/confusion.

Now, if you revert this check-in (afc2639) and then run check-mlir it would show 1 crash, the newly added test. With this above-mentioned check-in the whole thing should pass.

[dcaballe]

LG, thanks!

[bviyer]

Hi @joker-eph and @banach-space, Please let me know if you have any more questions. I would like to land the patch by Monday (March 25th 2024) evening if I don't hear any objections.

[banach-space]

I would like to land the patch by Monday (March 25th 2024) evening if I don't hear any objections.

Thanks for the update. I've made a suggestion how to improve the test - I'd like that to be addressed before this is merged.

[banach-space]

LGTM, thanks for addressing my comments!