Currently the V8 integration explicitly disables the Liftoff compiler. AIUI this is to force V8 to only use TurboFan for code generation, so that in uses such as Envoy where there are N separate thread-local VMs, we incur the cost of eager TurboFan code generation only once, prior to cloning each of the thread-local VMs.
However, Liftoff offers significant benefits for use cases requiring quick plugin startup time--in some benchmarks we have seen it consume <5% of the CPU time of TurboFan. For those use cases it would be useful to have the option to enable Liftoff for V8.
Currently the V8 integration explicitly disables the Liftoff compiler. AIUI this is to force V8 to only use TurboFan for code generation, so that in uses such as Envoy where there are N separate thread-local VMs, we incur the cost of eager TurboFan code generation only once, prior to cloning each of the thread-local VMs.
However, Liftoff offers significant benefits for use cases requiring quick plugin startup time--in some benchmarks we have seen it consume <5% of the CPU time of TurboFan. For those use cases it would be useful to have the option to enable Liftoff for V8.