RFR: 8253404: C2: assert(C->live_nodes() <= C->max_node_limit()) failed: Live Node limit exceeded limit

[Roberto Castañeda Lozano]

On Wed, 7 Oct 2020 11:42:44 GMT, Roberto Castañeda Lozano <github.com+8792647+robcasloz at openjdk.org> wrote:

>> Record nodes as dead in `Node::destruct()` if their index cannot be directly reclaimed. This prevents the "Live Node
>> limit exceeded limit" assertion failure by improving the accuracy of `Compile::live_nodes()` when "hook" nodes in
>> `ConvI2LNode::Ideal()` are created and deleted non-consecutively.  This addition might result in multiple calls to
>> `compile::record_dead_node()` for the same node (e.g. from `PhaseIdealLoop::spinup()`), but this is safe, as
>> `compile::record_dead_node()` is idempotent.
>
> Record nodes as dead in Node::destruct() if their index cannot be directly
> reclaimed. This prevents the "Live Node limit exceeded limit" assertion failure
> by improving the accuracy of Compile::live_nodes() when "hook" nodes in
> ConvI2LNode::Ideal() are created and deleted non-consecutively.
>     
> This addition might result in multiple calls to compile::record_dead_node() for
> the same node (e.g. from PhaseIdealLoop::spinup()), but this is safe, as
> compile::record_dead_node() is idempotent.

There is, potentially, a more fundamental issue with the exponential C2 execution time of the `ConvI2L(AddI(x, y)) ->
AddL(ConvI2L(x), ConvI2L(y))` optimization in `ConvI2LNode::Ideal()` when optimizing long chains of int add/sub nodes
followed by an int-to-long conversion, but I think that can be addressed separately in a RFE. The likelihood of hitting
this worst case in "real code" and the actual impact on total execution time is unclear.

-------------

PR: https://git.openjdk.java.net/jdk/pull/540