I've spent quite a few hours today trying to understand why this code segfaults on g++6.2
and g++7.0
, while happily working as intended on clang++3.9
(and 4.0
).
I reduced the issue to a 85 lines self-contained code snippet, which does not segfault upon normal execution, but always reports an error under UBSAN.
The issue is reproducible on wandbox, by compiling with g++7
, enabling optimizations and passing -fsanitize=undefined
as an extra flag.
This is what UBSAN reports:
prog.cc: In function 'int main()':
prog.cc:61:49: warning: 'ns#0' is used uninitialized in this function [-Wuninitialized]
([&] { ([&] { n.execute(ns...); })(); })();
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~
prog.cc:28:10: note: 'ns#0' was declared here
auto execute(TNode& n, TNodes&... ns)
^~~~~~~
prog.cc:30:9: runtime error: member call on null pointer of type 'struct node_then'
g++
claims that ns#0
is uninitialized inside the "lambda gibberish" (which simulates the for_tuple
from the original snippet). Now, some very interesting things occur:
If I remove the "lambda gibberish", transforming line 61 into
n.execute(ns...);
then UBSAN stops complaining.
If I change the capture list from
[&]
to[&n, &ns...]
, UBSAN stops complaining as well:([&](auto) { ([&n, &ns...] { n.execute(ns...); })(); })(0);
...wait what? How is that different from
[&]
?
Applying the above discoveries to the original code snippet fixes the segfaults.
Is this a g++
bug? Or is there any undefined behavior in my code?
This has nothing to do with temporaries: it's a gcc7.0 optimizer bug. This is a simpler reproducer:
Output:
Link: http://melpon.org/wandbox/permlink/E11fOumFJda6OW6m
To aid in this code's comprehension I'm using a powerful debugging tool:
paint.exe