smeenai added a comment.
Finally got a chance to dig into some of the failures this patch is causing.
Here's an example crasher (run with `clang -cc1 -triple i686-windows
-emit-llvm`):
struct Q {
Q(int);
~Q();
};
struct Z {};
struct A {
A(Q);
};
struct B : Z, A {
using A::A;
};
B b(Q(1));
What's happening here is that we emit the inlined inherited constructor call
(B's `using A::A`), which in turn emits the constructor call (`A::A`), which
eventually winds up in CodeGenFunction::EmitCall, which calls
deactivateArgCleanupsBeforeCall, which cleans the cleanup stack. However, the
RunCleanupsScope that I added thinks it has cleanups to do, because the
EHStack's stable_begin iterator holds a different (empty) value at its
destruction than it did at its construction. Of course, the cleanup stack is
actually empty at this point, so attempting to run cleanups just asserts.
I guess what I want is for my added scope to only apply to the cleanups added
for base destructor calls (since any cleanups for arguments will be handled by
deactivateArgCleanupsBeforeCall), but I'm not sure how best to accomplish that.
CodeGenFunction::EmitConstructorBody also wraps a RunCleanupsScope around its
EmitCtorPrologue call, so in theory it should be affected by the same problem.
I'll try to craft a test.
Repository:
rC Clang
https://reviews.llvm.org/D44619
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
