mark calls in the unwind path as !noinline

The unwind path is always cold, so that should not have bad performance
implications.  This avoids catastrophic exponential inlining, and also
decreases the size of librustc.so by 1.5% (OTOH, the size of `libstd.so`
increased by 0.5% for some reason).

Fixes #41696.

src/librustc_trans/mir/block.rs

@@ -146,6 +146,13 @@ fn trans_terminator(&mut self,
             } else {
                 let llret = bcx.call(fn_ptr, &llargs, cleanup_bundle);
                 fn_ty.apply_attrs_callsite(llret);
+                if this.mir[bb].is_cleanup {
+                    // Cleanup is always the cold path. Don't inline
+                    // drop glue. Also, when there is a deeply-nested
+                    // struct, there are "symmetry" issues that cause
+                    // exponential inlining - see issue #41696.
+                    llvm::Attribute::NoInline.apply_callsite(llvm::AttributePlace::Function, llret);
+                }
 
                 if let Some((ret_dest, ret_ty, target)) = destination {
                     let op = OperandRef {

src/test/run-pass/issue-41696.rs

@@ -9,8 +9,6 @@
 // except according to those terms.
 
 // this used to cause exponential code-size blowup during LLVM passes.
-// ignore-test FIXME #41696
-// min-llvm-version 3.9
 
 #![feature(test)]