RFR(S): 8029351: assert(bt != T_OBJECT) failed: Guard is incorrect in VM:defmeth

[Christian Thalinger]

It would only be used once but I think it would be a good idea to move the "is this an object?” test into a method in constantTag.  That way the checks would have a descriptive name and be co-located with the other tags.  Might help to avoid bugs like this in the future.

The fix looks good.

On Dec 7, 2013, at 7:12 PM, David Chase <david.r.chase at oracle.com> wrote:

> RT-dev CC'd because this was more of a runtime bug, except that involved methodHandles.
> 
> bug: https://bugs.openjdk.java.net/browse/JDK-8029351
> webrev: http://cr.openjdk.java.net/~drchase/8029351/webrev.01/
> 
> Running "defmeth", especially with -Xcomp and -XX:+AggressiveOpts, sometimes (1 in 25 to 1 in 50)
> causes a crash in fastdebug, that in release builds is instead a signal-ignoring infinite loop.
> 
> The cause was failure to include the method_handle_in_error and method_type_in_error
> cases in the "is this an object?" test.  This immediately led to a blown assertion in a debug build.
> Fix was to add the obvious clauses to the if:
> 
>       tag.is_unresolved_klass() ||
>       tag.is_string() ||
>       tag.is_method_handle() ||
> -      tag.is_method_type()) {
> +      tag.is_method_type() ||
> +      tag.is_method_handle_in_error() ||
> +      tag.is_method_type_in_error()) {
>     assert(bt == T_OBJECT, "Guard is incorrect");
>     cts = CellTypeState::make_line_ref(bci);
>   } else {
> 
> Testing:
> defmeth, fastdebug build with bug-provoking flags , 1000 times.
> All ran w/o crash (i.e., only with other existing known failures), all produced the same output.
> 
> David
>