1. Devirtualization of method calls
Dirty tricks of JIT compilers
Alexey Ragozin
alexey.ragozin@gmail.com
Apr 2013

2. Old good C++
010110010010
101010100110
101010100101
010101001010
101010101010
101010101010
101010101010
101010100010
00: methodA
02: methodC
03: methodD
CODEOBJECT
VTABLE
01: methodB
Simple case

3. Old good C++
010110010010
101010100110
101010100101
010101001010
101010101010
101010101010
101010101010
101010100010
111010100100
011110000010
101001010100
00: methodA
02: methodC
03: methodD
CODEOBJECT
VTABLE
01: methodB
00: methodX
02: methodZ
01: methodY
VTABLE
Multiple inheritance

4. Old good C++
Multiple inheritance – more fun
A
B C
D D
A
B C
D
Vs

5. Problems
Two memory reads before jump
 Memory access is serialized
 instruction pipeline is blocked

6. Dynamic languages
 Every call is virtual
 No predefined classes
 Multiple inheritance
Nightmare for super scalar CPU

7. Call site polymorphism
A particular call site
in most case
would be calling
exactly one method instance
Call sites are mostly monomorthic!

8. Exploiting monomorthism
Tracing JIT
 JavaScript – Mozila TraceMonkey
 Python – PyPy
Profiling for whole method JIT
 JVM

9. Tracing JIT
Interpretation phase
 Record actions, record branch conditions
Compiling trace
 Actions compiled into branchless machine code
 Guards added for conditions
 If guard is broken – fall back to interpretation

10. Tracing JIT
Mozila’s TraceMonkey
1. IncrementalDynamicCodeGenerationwithTraceTrees
http://www.ics.uci.edu/~franz/Site/pubs-pdf/ICS-TR-06-16.pdf
PyPy
 RPython – restrict python dialect compiling to C
 Rpython based interpreter trace JITed by runtime
1. http://tratt.net/laurie/research/pubs/papers/bolz_tratt__the_impact_of_metatra
cing_on_vm_design_and_implementation.pdf

11. Tracing JIT
Limitation
 Trace is EXPENSIVE
 Falling out of path is EXPENSIVE
 A lot of code is executed just once
– wasting resources on tracing

12. Whole method JIT
JVM
 Classic whole method JIT
 Profile “morphism” of call sites
 Multi tier compiler
 Dynamics recompilation / decompilation

13. Whole method JIT
Call sites
 Monomorphic
 Bimorphic
 Megamorphic

14. Invoke dynamic
Problem
 Interpreters tends to produce
megamorphic sites
Invoke dynamic
 Detach call site into manageable object

15. More JVM back magic
On stack replacement
 Code for method could be replaced without
reentering to a method
Scalar replacement
 Object not leaving method scope
can be reduced to a number of local variables

16. THANK YOU
alexey.ragozin@gmail.com