Created attachment 15482 [details] Testcase The Schedule generation relies on a reverse post order traversal as the domain generation does. Though, for the former this is not the right choice as it does not guarantee loops are visited completely after the header was visited. Instead, blocks after the loop can be visited first causing them to appear prior in the schedule than they should. An example test case is attached. This bug was detected due to a SSA-Codegen failure in an lnt benchmark. It does apperently not cause a result change for the current code generation, even though we basically load/use an uninitialized value... I am working on a patch for the schedule generation but ideas are welcome.
Move bugs to Polly product.
https://llvm.org/svn/llvm-project/polly/trunk@259354