================
@@ -325,19 +374,47 @@ mlir::LogicalResult
ItaniumEHLowering::lowerFunc(cir::FuncOp funcOp) {
if (!funcOp.getPersonality())
funcOp.setPersonality(kGxxPersonality);
- // Lower each initiate and all EH ops connected to it. The token map is
- // shared across all initiate operations. Multiple initiates may flow into
the
- // same dispatch block, and the map ensures the arguments are registered
- // only once. Dispatch ops are scheduled for deferred removal so that sibling
- // initiates can still read catch types from a shared dispatch.
+ // Compute, read-only and before any destructive lowering, the dispatches
each
+ // initiate's exception can reach (innermost first; more than one for nested
+ // try/catch). A landing pad's catch types are a property of the EH graph,
so
+ // deriving them here keeps them independent of the order in which the
+ // destructive per-initiate traversal in lowerEhInitiate tears down shared
+ // token-graph edges. Otherwise a sibling or outer dispatch could be missed,
+ // leaving a landing pad without its catch clause (it would resume past the
+ // handler to std::terminate) or an un-lowered leftover dispatch.
+ llvm::DenseMap<mlir::Operation *, SmallVector<cir::EhDispatchOp>>
+ reachedDispatches;
+ llvm::DenseMap<mlir::Operation *, bool> reachesCleanup;
----------------
andykaylor wrote:
Could you combine these two maps? It looks like everywhere that you lookup an
entry in one, you immediately look up an entry in the other with the same key.
https://github.com/llvm/llvm-project/pull/205638
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
