The Cog VM evolution Video: https://youtu.be/Gi0WiIxvJ60 Thu, August 25, 11:30am – 12:00pm First Name: Clement Last Name: Bera Email: bera.clement@gmail.com Title: The Cog VM evolution Type: Talk Abstract: In 2008, the Squeak VM repository made by Dan Ingalls was forked to give birth to the high-performance Cog VM, which is now the default VM for multiple Smalltalk dialects such as Pharo, Squeak or Cuis. Since the fork, new features and performance enhancements have been introduced every year. This talk sums up all the main features and main performance tweaks that were added to the Cog VM in chronological order from 2008 to today. The focus is on the execution engine: Memory management, interpretation, JIT compilation and FFI. The talk lastly discusses the current work in-progress and short-term future work. Bio: Clément Béra is a PhD student working on virtual machines for object oriented languages at Inria, Lille. He is now focusing on runtime optimisations for the Cog virtual machine and its Smalltalk clients.

1. CogVM
Evolution
Clément Béra
Thursday, August 25, 16

2. CogVM ?
• Smalltalk virtual machine
• DefaultVM for
• Pharo
• Squeak
• Newspeak
• Cuis
Thursday, August 25, 16

3. Cog Philosophy
• Open source (MIT)
• Simple
• Is the optimization / feature worth the
added complexity ?
• Cross-Platform (Processors, OS, 32/64 bits)
Thursday, August 25, 16

4. Execution engine
• VM
• Execution engine
• Plugins: graphics, file, etc.
Thursday, August 25, 16

5. Execution engine
• Interpreter
• JIT
• Memory Manager (including GC)
Thursday, August 25, 16

6. Evolution
• User and Customer driven
• Where did we start ?
• What problems did we solve ?
Thursday, August 25, 16

7. Starting blocks
• InterpreterVM
• Made by Dan Ingalls Team
• Simple Interpreter
• Spaghetti stack
• Smart but simple Memory Manager
Thursday, August 25, 16

8. Performance !
Short-term delivery
Performance for 3D application
Thursday, August 25, 16

9. Performance !
Short-term delivery
Performance for 3D application
• Fast Interpreter
Thursday, August 25, 16

10. • Context-to-stack Mapping
• 85% of context allocation removed
• No copying of arguments
• New hash logic
• Primitive function caching
StackVM
Thursday, August 25, 16

11. Issue
• Closure implementation
• No temporaries
• BlockAlreadyEvaluated error
• Non obvious bugs
• New implementation
Thursday, August 25, 16

12. 0
2
4
6
8
10
12
14
Interpreter Stack Cog V1 Cog V2 Spur Sista
Binary Tree benchmark
Thursday, August 25, 16

13. More Performance !
Short-term delivery
Performance for 3D application
Thursday, August 25, 16

14. More Performance !
Short-term delivery
Performance for 3D application
• First JIT compiler
Thursday, August 25, 16

15. CogVM
• x86 back-end
• Simple machine code generation
• Except inline caches
Thursday, August 25, 16

16. Binary Tree benchmark
0
2
4
6
8
10
12
14
Interpreter Stack Cog V1 Cog V2 Spur Sista
Thursday, August 25, 16

17. JIT Abstractions
V3
Object
Representation
SimpleStackCogit
Cogit
Implementation
x86
Machine
back-end
Thursday, August 25, 16

18. Yet More Performance !
Short-term delivery
Performance for 3D application
Thursday, August 25, 16

19. Yet More Performance !
Short-term delivery
Performance for 3D application
• Second JIT compiler
Thursday, August 25, 16

20. CogVM
• Machine code generation
• Linear scan register allocation
• Avoids many stack operations
• Register-based calling convention
Thursday, August 25, 16

21. Binary Tree benchmark
0
2
4
6
8
10
12
14
Interpreter Stack Cog V1 Cog V2 Spur Sista
Thursday, August 25, 16

22. JIT Abstractions
V3
Object
Representation
SimpleStackCogit
StackToRegisterMappingCogit
Cogit
Implementation
x86
Machine
back-end
Thursday, August 25, 16

23. Newspeak support ?
Newspeak is Smalltalk-like
New kind of sends
Thursday, August 25, 16

24. Newspeak support ?
• Multiple bytecode set support
• Newspeak specific operations
• Interpreter
• Machine code generation
Thursday, August 25, 16

25. 64 bits ?
Images larger than 1 or 2 Gb ?
Moving objects during FFI call-backs ?
Even more performance !
Ephemerons ?
Become is so slow it cannot be used.
Thursday, August 25, 16

26. 64 bits ?
Images larger than 1 or 2 Gb ?
Moving objects during FFI call-backs ?
Even more performance !
Ephemerons ?
Become is so slow it cannot be used.
Short-term delivery
Thursday, August 25, 16

27. 64 bits ?
Images larger than 1 or 2 Gb ?
Moving objects during FFI call-backs ?
Even more performance !
Ephemerons ?
Become is so slow it cannot be used.
Short-term delivery
• New Memory Manager
Thursday, August 25, 16

28. Spur Memory Manager
• Class-table (efficient caches and compactness)
• Efficient scavenging
• Fast become
• New object layouts (Ephemerons, ShortArrays)
• Memory representation 64-bits compatible
• Pinned objects
• Segmented Memory
Thursday, August 25, 16

29. Binary Tree benchmark
0
2
4
6
8
10
12
14
Interpreter Stack Cog V1 Cog V2 Spur Sista
Thursday, August 25, 16

30. JIT Abstractions
V3
Spur32
Spur64
Object
Representation
SimpleStackCogit
StackToRegisterMappingCogit
Cogit
Implementation
x86
Machine
back-end
V3
Spur32
Spur64
Object
Representation
x86
Machine
back-end
32 bits 64 bits
Thursday, August 25, 16

31. Raspberry Pi performance ?
Scratch support
Thursday, August 25, 16

32. • ARMv6 support
Raspberry Pi performance ?
Scratch support
Thursday, August 25, 16

33. ARMv6 support
• JIT ARMv6 back-end
• JIT abstraction over literal management
• JIT abstraction over CISC / RISC
Thursday, August 25, 16

34. JIT Abstractions
V3
Spur32
Spur64
Object
Representation
SimpleStackCogit
StackToRegisterMappingCogit
Cogit
Implementation
x86
Machine
back-end
V3
Spur32
Spur64
Object
Representation
x86
ARMv6
MIPSEL
Machine
back-end
32 bits 64 bits
Inline
Outline
Literal
Manager
Thursday, August 25, 16

35. Ryan (contributor)
Working with the DartVM
Dart runs on more platform than Newspeak.
Thursday, August 25, 16

36. Ryan (contributor)
Working with the DartVM
Dart runs on more platform than Newspeak.
• MIPSEL support
Thursday, August 25, 16

37. JIT Abstractions
V3
Spur32
Spur64
Object
Representation
SimpleStackCogit
StackToRegisterMappingCogit
Cogit
Implementation
x86
Machine
back-end
V3
Spur32
Spur64
Object
Representation
x86
ARMv6
MIPSEL
Machine
back-end
32 bits 64 bits
Inline
Outline
Literal
Manager
Thursday, August 25, 16

38. 64 bits support ?
64 bits library binding
Heap over 2Gb
Thursday, August 25, 16

39. 64 bits support ?
64 bits library binding
Heap over 2Gb
• x64 support
• Immediate float
Thursday, August 25, 16

40. JIT Abstractions
V3
Spur32
Spur64
Object
Representation
Inline
Outline
Literal
Manager
SimpleStackCogit
StackToRegisterMappingCogit
Cogit
Implementation
x86
ARMv6
MIPSEL
x64
Machine
back-end
32 bits 64 bits
Thursday, August 25, 16

41. Yet even more performance !
Computation lasting
3 to 6 hours
Thursday, August 25, 16

42. Yet even more performance !
Computation lasting
3 to 6 hours
• Speculative optimizations
Thursday, August 25, 16

43. SistaVM
• The program introspects
• Optimize the code for performance
• Deoptimize when it took incorrect decisions
Thursday, August 25, 16

44. Issues
• Bytecode set
• Literal mutability
• Closure implementation
Thursday, August 25, 16

45. Bytecode set limitations ?
Code generator
tools
Thursday, August 25, 16

46. Sista bytecode set
• Lifting encoding limitations
• Encode instructions for the Sista / Lowcode
Thursday, August 25, 16

47. Efficient modification trackers ?
Literal inconsistency ?
Thursday, August 25, 16

48. • Read-only objects
• IWST talk this afternoon
• Hopefully allows more compiler optimizations
Efficient modification trackers ?
Literal inconsistency ?
Thursday, August 25, 16

49. Closure implementation
• Method and closure get more similar
• Simplifies part of theVM
• Simplifies the runtime compiler
Thursday, August 25, 16

50. Integration
• Closed alpha version
• Integrating dependencies:
• Bytecode set integrated
• Read-only objects integrated (but disabled)
• Closure implementation in progress
Thursday, August 25, 16

51. Binary Tree benchmark
0
2
4
6
8
10
12
14
Interpreter Stack Cog V1 Cog V2 Spur Sista
Thursday, August 25, 16

52. JIT Abstractions
V3
Spur32
Spur64
Object
Representation
Inline
Outline
Literal
Manager
SimpleStackCogit
StackToRegisterMappingCogit
RegisterAllocatingCogit
SistaCogit
Cogit
Implementation
x86
ARMv6
MIPSEL
x64
Machine
back-end
32 bits 64 bits
Thursday, August 25, 16

53. Image compaction ?
Sometimes, large images
when saving
Thursday, August 25, 16

54. Image compaction ?
Sometimes, large images
when saving
• Better compactor
Thursday, August 25, 16

55. Pauses ?
0.5 second freezes
in UI application
Thursday, August 25, 16

56. Pauses ?
0.5 second freezes
in UI application
• Incremental GC
Thursday, August 25, 16

57. Many hidden parts
...
Thursday, August 25, 16

58. • C generated from Slang
• Many were fixed
• Towards compilation with -WAll -WError
C compiler warning ?
Thursday, August 25, 16

59. • LargeInteger plugin more efficient
• Computation moved from 8bits to 32 bits
• Different compilation flags
Faster arithmetic ?
Thursday, August 25, 16

60. • Slang-to-C compiler
• Many improvements
• Type inference
Slang ?
Thursday, August 25, 16

61. Contribution
Thursday, August 25, 16

62. CogVM team
• Started with Eliot Miranda
• Many more contributors now:
• Tim Rowledge
• Clément Béra (myself)
• Nicolas Cellier
• Fabio Niephaus & Tim Felgentreff
• Ryan Macnak
Thursday, August 25, 16

63. Conclusion
• Lots of new features and improvements over years
• A lot more is incoming
• If you want to support, talk to us !
• ARMv8 ?
• Incremental GC ?
• Performance (escaping, floats) ?
Thursday, August 25, 16