Pin panic-in-drop=abort test to old pass manager

1 files changed, 8 insertions, 1 deletions

diff --git a/src/test/codegen/panic-in-drop-abort.rs b/src/test/codegen/panic-in-drop-abort.rs
index 0d19351a33d..39f73c4e396 100644
--- a/src/test/codegen/panic-in-drop-abort.rs
+++ b/src/test/codegen/panic-in-drop-abort.rs
@@ -1,8 +1,15 @@
-// compile-flags: -Z panic-in-drop=abort -O
+// compile-flags: -Z panic-in-drop=abort -O -Z new-llvm-pass-manager=no
 
 // Ensure that unwinding code paths are eliminated from the output after
 // optimization.
 
+// This test uses -Z new-llvm-pass-manager=no, because the expected optimization does not happen
+// on targets using SEH exceptions (i.e. MSVC) anymore. The core issue is that Rust promises that
+// the drop_in_place() function can't unwind, but implements it in a way that *can*, because we
+// currently go out of our way to allow longjmps, which also use the unwinding mechanism on MSVC
+// targets. We should either forbid longjmps, or not assume nounwind, making this optimization
+// incompatible with the current behavior of running cleanuppads on longjmp unwinding.
+
 // CHECK-NOT: {{(call|invoke).*}}should_not_appear_in_output
 
 #![crate_type = "lib"]