[llvm] [mlir] [Utils][mlir] Fix interaction between CodeExtractor and OpenMPIRBuilder (PR #145051)

[Tom Eccles via llvm-commits]

================
@@ -1868,8 +1866,14 @@ CallInst *CodeExtractor::emitReplacerCall(
   // This takes place of the original loop
   BasicBlock *codeReplacer =
       BasicBlock::Create(Context, "codeRepl", oldFunction, ReplIP);
+  // In cases with multiple levels of outlining, e.g. with OpenMP,
+  // AllocationBlock may end up in a function different than oldFunction. We
+  // need to make sure we do not use it in those cases, otherwise the alloca
+  // will end up in a different function from its users and break the module.
   BasicBlock *AllocaBlock =
-      AllocationBlock ? AllocationBlock : &oldFunction->getEntryBlock();
+      (AllocationBlock && oldFunction == AllocationBlock->getParent())
+          ? AllocationBlock
+          : &oldFunction->getEntryBlock();
----------------
tblah wrote:

Thanks for the update.

Sorry to be difficult but looking at this change gave me an idea: I think this situation just shouldn't ever happen.

Say we have something like this
```
omp.op1 {
  other.op1
  omp.op2 {
    other.op2
  }
  other.op3
}
```

If the lowering for omp.op2 is trying to put allocas in the alloca region for op1 then yes both of those could be outlined to different functions. But I don't think this should ever happen. If omp.op2 is expecting to be outlined, I think it should define its own alloca insertion point on the alloca stack.

I suspect one of the operations in your test needs to add an insertion point to the stack.

https://github.com/llvm/llvm-project/pull/145051