Request for review (XS): 6901572 JVM 1.6.16 crash on loops: assert(has_node(i),"")

[Vladimir Kozlov]

Changpeng,

Update the comment in loopnode.cpp for dead phi check.

So we hit the assert in build_loop_late because in
the secondary IV code we added new users of the phi
and set control for them (phase->set_ctrl(add, cl)).
Is this assumption correct? Or it was different reason?

Vladimir

Changpeng Fang wrote:
> http://cr.openjdk.java.net/~cfang/6901572/webrev.01/
> 
> Fixes 6901572 JVM 1.6.16 crash on loops: assert(has_node(i),"")
> 
> Problem:
> When the loop optimizer handles secondary induction variables in 
> IdealLoopTree::counted_loop,
> it hits a dead phi and results in "assert(has_node(i),"")" at the time 
> of  "get_ctrl" for the dead node.
> The phi is dead because "final int x" in the attached test case is 
> useless and becomes dead after the
> safepoint is removed for the counted loop.
> 
> Solution:
> Skip the secondary induction variable handling if it is dead. The dead 
> phi will be put into the deadlist
> late in PhaseIdealLoop::build_loop_late and removed.
> 
> Test:
> test case in the bug report.
> 
> Thanks,
> 
> Changpeng