[Valgrind-developers] How can I reproduce this behavior with some simple source codes ?

[邓尧 <to...@gm...>]

Hi,
I'm running into a very weird valgrind behavior, not sure whether it's a
bug in my code or bug in valgrind.

The following is source code that runs well:

void ServerPool::processConnection(int connection) {
int offset = getWorker();
Context& context = pool->getContext(offset);
int* fd = &context.pipe[1];
wrapper::write_(*fd, &connection, sizeof(connection));
lastWorker = offset;
}

Context is defined as:

struct Context {
ServerPool* volatile self;
int pipe[2];

Context();
virtual ~Context();
};

If the following two lines:
int* fd = &context.pipe[1];
wrapper::write_(*fd, &connection, sizeof(connection));
with
int fd = context.pipe[1];
wrapper::write_(fd, &connection, sizeof(connection));
the program will crash, I checked the coredump generated by valgrind, in
the value of variable "fd" is
-1. Both code would run correctly without valgrind.

I got this situation under both valgrind 3.7.0 and 3.8.1, ubuntu linux
12.04 64-bit. source code compiled with gcc 4.6.3, glibc version 2.15

BTW, wrapper::write_() is simple a wrapper function around the system call
write(), it checks the return value of write(), if the value is less than
expected, it will throw an exception. pool->getContext() simply returns a
reference to a new'ed instance.

Thanks
Yao