Details
-
Sub-task
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
None
-
Patch Available
-
Moderate
Description
Now, client JIT chains look like this:
chains=chain1,chain2
chain1.jits=JET_CLINIT
chain2.jits=JET_DPGO,CD_OPT
chain1.filter=+::<clinit>
chain1.filter=-
JET_CLINIT.file=jitrino
JET_DPGO.file=jitrino
CD_OPT.file=jitrino
JET_DPGO.genProfile=EB_PROF
EB_PROF.profilerType=EB_PROFILER
CD_OPT.useProfile=EB_PROF
So, all methods (except class initializers) are passing through JET_DPGO with EB_PROF attached and then will be recompiled by CD_OPT. One should note that this will cause concurrency between on-going recompilation and first-time JET compilation. So, if first-time JET compilation speed is important for startup, then we should separate these two processes in time. That is, let's compile all in the bunch with JET and only then start to recompile with OPT.
That is, we can either:
a. Set the thresholds for EB_PROF such as recompilation would held later
b. Introduce new step (e.g. JET_QUICK) which would compile really fast (even without profile) and wait before inducing recompilation.
Current architecture of Jitrino requires some profiler to do the notification for traversing the JITs chain, so attached patch introduces dummy DELAY_PROFILER which simply waits desired time.
So, we will have:
chains=chain1,chain2
chain1.jits=JET_CLINIT
chain2.jits=JET_QUICK,JET_DPGO,CD_OPT
chain1.filter=+::<clinit>
chain1.filter=-
JET_CLINIT.file=jitrino
JET_QUICK.file=jitrino
JET_DPGO.file=jitrino
CD_OPT.file=jitrino
JET_DPGO.genProfile=EB_PROF
EB_PROF.profilerType=EB_PROFILER
CD_OPT.useProfile=EB_PROF
JET_QUICK.genProfile=DL_PROF
DL_PROF.profilerType=DL_PROFILER
DL_PROF.initialTimeout=0
DL_PROF.delayTimeout=100