VCpu is a framework for writing low-level code and simulating or generating machine code for backends like ARM and x86. It allows coding directly in Smalltalk by representing variables as machine words and supporting arithmetic, memory access, control flow, and calling conventions. While not a full compiler, VCpu provides a simple yet powerful way to express low-level code and interface with libraries like NativeBoost without knowing multiple disciplines.

1. Virtual CPU
ESUG, Cambridge 2014
By Igor Stasenko & Max Mattone
RMod, Inria
Wednesday, August 20, 14

2. Highlights
• What?
• Why?
• How?
• Demo
• To Do
Wednesday, August 20, 14

3. What is VCpu?
• a framework to write low-level code
• can simulate & generate machine code
• multiple backends for ARM, x86/x64
code generation
• 100% implemented in smalltalk
Wednesday, August 20, 14

4. What is VCpu NOT
• NOT a full-fledged compiler with
numerous data “types”, like GCC/
LLVM
• NO direct support of calling
convention(s)
• it is a bare-bone model of computer
with CPU & memory.. to build on top of
it
Wednesday, August 20, 14

5. Requirements
• expressive power of smalltalk
• malleable
• extensible
• simple
• yet powerful
Wednesday, August 20, 14

6. Architecture
Platform neutral
RTL VCPU Interface
Simulating CPU Generating CPU
Low-level intermediate
Liveness analysis
Register allocation
Optimizations
AsmJIT
Platform dependent
Wednesday, August 20, 14

7. Why
• there’s no low-level compilers for
smalltalk
• on inventing own wheel: adapting
existing solutions costly as (h/w)ell
• lets us learn as we do it
Wednesday, August 20, 14

8. Don’t mode me in
Implementing low-level semantic
Idea Smalltalk Slang C Machine
Idea Machine
Wednesday, August 20, 14

9. A more correct picture
Idea Machine
Wednesday, August 20, 14

10. At the end of the day
memory at: x put: y
So, tell me, why you have to be expert in 10+
disciplines to do that?
Wednesday, August 20, 14

11. There must be a better
way
• i was looking for a nice & simple
solution since 2006
• VCpu interface, is product of number
of painful attempts to get there
• Now i am pleased (almost).
Wednesday, August 20, 14

12. How
• imperative rather than declarative
• coding with VCpu is just plain
smalltalk
Wednesday, August 20, 14

13. Imperative
• you don’t ‘compile’ or translate code,
you just execute own code istructing
CPU what to do: cpu doThat
Wednesday, August 20, 14

14. Dual nature
• can be either simulated or generating
machine code, just use different CPU
• you free to choose any style you want
to program it
Wednesday, August 20, 14

15. Machine word
• a facade object representing a virtual
CPU register/variable (machine word)
• it easy to manipulate with, since one
can define a usual arithmetic
operations, like #+ , #- , #*, #/ etc..
• serves as a basis of VCpu ‘DSL’
Wednesday, August 20, 14

16. Lets learn a new DSL
Wednesday, August 20, 14

17. Step 1. Creating a new
machine word
word := cpu word: 10.
Wednesday, August 20, 14

18. Step 2. Assigning new
value to existing one
word1 := cpu word: 10.
word2 := cpu word: 4.
...
word2 value: word1.
Right
word2 := word1. Wrong!!!
Wednesday, August 20, 14

19. Step 3. Arithmetic
expressions
word := x + y bitAnd: z
just keep in mind, it is not ‘school’ but ‘CPU’ math
Wednesday, August 20, 14

20. Step 3. Arithmetic
expressions
the result of expression is
always new machine word
word := x + y bitAnd: z
just keep in mind, maybe you wanted:
word value: x + y bitAnd: z
... instead
Wednesday, August 20, 14

21. Step 3. Arithmetic
expressions
.. expressions can be intermixed with regular
constants
word := x + 5
as long as receiver is machine word
Wednesday, August 20, 14

22. Step 4. Memory access
word := address loadWord
address writeWord: x
Wednesday, August 20, 14

23. Step 5. Comparisons/
control flow
a>b ifTrue: [..] ifFalse: [..]
a to: b do: [:i |..]
x timesRepeat: [ .. ]
Looks familiar?
Wednesday, August 20, 14

24. Thanks, we have Opal
• disable inlining (ifTrue:/to:do: ..etc)
expression ifTrue: [..] ifFalse: [..]
mustBeABoolean
this is Sparta Pharo!
Wednesday, August 20, 14

25. Step 6. Call/return
address call
cpu return.
cpu return: x **
** requires a notion of calling convention
Wednesday, August 20, 14

26. End of tutorial
Wednesday, August 20, 14

27. NativeBoost integration
abs: x
NativeBoost-style callout
<primitive: #primitiveNativeCall module: #NativeBoostPlugin>
^ self nbCallout
function: #( int abs (int x ) )
module: NativeBoost CLibrary
VCpu-style callout
?
Wednesday, August 20, 14

28. NativeBoost integration
abs: x
NativeBoost-style callout
<primitive: #primitiveNativeCall module: #NativeBoostPlugin>
^ self nbCallout
function: #( int abs (int x ) )
module: NativeBoost CLibrary
abs: x
VCpu-style callout
<primitive: #primitiveNativeCall module: #NativeBoostPlugin>
^ self nbCallout
function: #( int abs (int x ) )
module: NativeBoost CLibrary
this is Sparta Pharo!
Wednesday, August 20, 14

29. Demo
Wednesday, August 20, 14

30. To Do
• Finish optimizations
• Test ARM Support (on real hardware)
• Complete NativeBoost VCpu
implementation
• Documentation
• Look forward for Spur integration
Wednesday, August 20, 14

31. ?
Wednesday, August 20, 14