[clang] [CIR] Fix cir.cleanup.scope successor regions (PR #221222)

[Henrich Lauko via cfe-commits]

================
@@ -2037,14 +2037,31 @@ LogicalResult cir::CleanupScopeOp::verify() {
 
 void cir::CleanupScopeOp::getSuccessorRegions(
     mlir::RegionBranchPoint point, SmallVectorImpl<RegionSuccessor> &regions) {
-  if (!point.isParent()) {
-    regions.emplace_back(getOperation());
+  // Execution always starts in the body region.
+  if (point.isParent()) {
+    regions.emplace_back(&getBodyRegion());
     return;
   }
 
-  // Execution always proceeds from the body region to the cleanup region.
-  regions.push_back(RegionSuccessor(&getBodyRegion()));
-  regions.push_back(RegionSuccessor(&getCleanupRegion()));
+  mlir::Operation *term = point.getTerminatorPredecessorOrNull();
+  assert(term && "expected a terminator predecessor");
+
+  if (term->getParentRegion() == &getBodyRegion()) {
----------------
xlauko wrote:

No. `cir.unreachable` and `cir.trap` don’t implement `RegionBranchTerminatorOpInterface`, so they aren’t branch points and get no outgoing edges. That’s right, since control never leaves through them and no cleanup runs.

Coroutine terminators should be handled correctly.

https://github.com/llvm/llvm-project/pull/221222