ZJIT: Function Lifecycle

[k0kubun]

Purpose

• Design a state machine for function lifecycle that maximizes the performance of ZJIT.
• Every state transition of a function should have a tuning knob that can be tuned easily.

Considerations

Possible tuning knobs to consider when designing function lifecycle:

• State transition:
  • How many times should we profile (or run non-profiling cycles) before compiling a version?
  • What kind of events (certain kinds of exits or invalidations) should trigger recompilation?
• Versioning:
  • How many versions should we compile for each function?
  • What compilation policy should we use for each version?
• Profiling
  • Should we have a profiling tier in JIT?
    • Can it profile something that the interpreter can't profile?
    • Would we afford to collect more profiles if we profiled faster?
    • How much extra memory do we need to generate JIT code for profiling?
  • Are there anything we aren't profiling but is profitable for JIT?
    • Branch probability
    • Profiles on a megamorphic ISEQ that is inlined by multiple JIT functions

TODO

• Research what kind of design decisions existing compilers made for the above design space and why.
• Make sure we have a tuning knob for everything we're interested in tuning.
  • Are function stubs compiling the callee immediately, not running enough profiling cycles?
  • Do we run the intended amount of cycles after resetting the call counter on invalidation or a recompile exit?
• Attempt to encapsulate the state machine in a single module, which makes it easier to trace, understand, and modify that.
• Make sure we don't miss optimization opportunities due to no profiles or compiling too early in benchmarks and SFR/Core.
  • Where are we seeing no-profiles right now?