On 27/11/2015 3:24 AM, Andrew Haley wrote:
I've been getting a lot of warnings such as
warning: 'size' may be used uninitialized in this function
[-Wmaybe-uninitialized]
which error out with -Werror. Almost all of them are bogus. They are
typically of the form
unsigned size;
if (i->get(26, 26)) { // float
switch(i->get(31, 30)) {
case 0b10:
size = 2; break;
case 0b01:
size = 1; break;
case 0b00:
size = 0; break;
default:
ShouldNotReachHere();
}
} else {
size = i->get(31, 31);
}
The problem here is that GCC does not know that ShouldNotReachHere()
should be treated as an unreachable statement.
Strictly speaking it is of course reachable, but if we do reach it we
expect never to return. As per the thread Mario pointed to we ran into
problems trying to mark this as not returning. But I wonder whether
lying to the compiler about the reachability of it would be a
workaround? Of course if the compiler used that information to elide the
ShouldNotReachHere() then that is not acceptable.
David
The patch here fixes it. I'd rather do this than add pointless assignments
all over the place. Thoughts? Opinions?
Thanks,
Andrew.
diff --git a/src/share/vm/utilities/debug.hpp b/src/share/vm/utilities/debug.hpp
--- a/src/share/vm/utilities/debug.hpp
+++ b/src/share/vm/utilities/debug.hpp
@@ -172,16 +172,24 @@
BREAKPOINT;
\
} while (0)
+#ifdef __GNUC__
+# define UNREACHABLE __builtin_unreachable()
+#else
+# define UNREACHABLE do { } while (0)
+#endif
+
#define ShouldNotReachHere()
\
do {
\
report_should_not_reach_here(__FILE__, __LINE__);
\
BREAKPOINT;
\
+ UNREACHABLE;
\
} while (0)
#define Unimplemented()
\
do {
\
report_unimplemented(__FILE__, __LINE__);
\
BREAKPOINT;
\
+ UNREACHABLE;
\
} while (0)
#define Untested(msg)
