RFR(M): 8031752: Failed speculative optimization should be reattempted when root of compilation is different

Roland Westrelin roland.westrelin at oracle.com
Wed Jan 15 03:19:54 PST 2014


When C2 optimizes the code based on type profiling, it emits a guard. If the guard fails, the code "traps" and the location of the trap (method, bci) and cause (class check) are recorded. When C2 generates code for that method again, it doesn't perform any optimization of that type (class check) at that location (bci). 

If, say, the trap occurs at (m, bci) when m is inlined in m1. When m2 that inlines m is compiled, no class check optimization is performed at (m, bci) because of the trap triggered in m1. With type speculation, type information is flowing from many profiling points in the application code. So in the example above, the profile data used at (m, bci) may come from m1 when the compiled method is m1 and it may be come from m2 when the compiled method is m2. And so the trap at (m,bci) in compiled method m1 doesn't say much about the validity of a speculative optimization at (m,bci) when the compiled method is m2. 

When a trap occurs from a speculative optimization, the trap should record: (method, bci, compiled method) so that in the example above: erroneous optimization at (m, bci) is not reattempted if m1 is re-compiled but is attempted when m2 is compiled.

With nashorn, peak performance varies a lot from one run to another for some benchmarks. This is one of the causes. Depending on the order of compilations, what’s compiled or not, what’s inlined or not, some optimizations may or may not be performed.

This change adds a new type of traps (SpeculativeTrapData). They record the nmethod’s method where the trap occurs. They are allocated in the extra data space of the MDO. If no more space is available in the extra data space, the VM fallbacks to standard traps.

I also added a PerMethodSpecTrapLimit similar to PerMethodTrapLimit but for speculative traps. Its value is much higher. That appears to be required as well to have reproducible peak performance results from one run to another.

I have a couple more changes that help reduce performance variation with nashorn and that I will send for review later.


More information about the hotspot-compiler-dev mailing list