[mlir] fix affine-loop-fusion crash

[lipracer]

If user not lies in Region findAncestorOpInRegion will return nullptr.

Fixes #76281.

[llvmbot]

@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-mlir-affine

Author: long.chen (lipracer)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/76351.diff

1 Files Affected:

• (modified) mlir/lib/Dialect/Affine/Transforms/LoopFusion.cpp (+2-1)

diff --git a/mlir/lib/Dialect/Affine/Transforms/LoopFusion.cpp b/mlir/lib/Dialect/Affine/Transforms/LoopFusion.cpp
index 66d921b4889f59..2140ff66b1a676 100644
--- a/mlir/lib/Dialect/Affine/Transforms/LoopFusion.cpp
+++ b/mlir/lib/Dialect/Affine/Transforms/LoopFusion.cpp
@@ -205,7 +205,8 @@ static bool isEscapingMemref(Value memref, Block *block) {
   // (e.g., call ops, alias creating ops, etc.).
   return llvm::any_of(memref.getUsers(), [&](Operation *user) {
     // Ignore users outside of `block`.
-    if (block->getParent()->findAncestorOpInRegion(*user)->getBlock() != block)
+    auto ancestorOp = block->getParent()->findAncestorOpInRegion(*user);
+    if (!ancestorOp || ancestorOp->getBlock() != block)
       return false;
     return !isa<AffineMapAccessInterface>(*user);
   });

[rikhuijzer]

Can you add a test which shows that #76281 is fixed? Preferably, also simplify the code from #76281 a bit so that the test is relatively simple. Without such a test, we might need to fix this issue again in the future.

Also, add fixes #76281 to your description in the first message on this PR. That will ensure that the linked issue is closed automatically when this PR is merged.

[rikhuijzer]

Apart from the suggestion and assuming the buildkite tests pass, this LGTM. Thanks for fixing the bug.

[rikhuijzer]

I've added Fixes #76281 to the first message so that this PR is linked to #76281. Once this PR is merged, #76281 will automatically be closed (https://docs.github.com/en/issues/tracking-your-work-with-issues/linking-a-pull-request-to-an-issue).