VM improvement: speeding up failing calls of functions in missing modules

Tim Rowledge tim at sumeru.stanford.edu
Sat Feb 5 18:52:48 UTC 2000



Stephan Rudlof wrote:

> If I understand you correctly the RTOSmalltalk codegenerator compiles
> directly from ST to Assembler without intermediate C code. This is
> interesting, but leads to some questions:
> 
> How difficult is it to call the OS by ANSI-C functions?
Not at all - you need to make the last stage object code generator
understand the OS calling conventions. It would be very like the glue
code that Ian's jitter uses for the same purpose

> How difficult is it to ensure compatibility between different platforms?
Not sure. The design was intended to be quite flexible and have
pluggable back ends for different CPUs. It builds a parse tree and then
generates code from that; curently I think C, ARM assembler source or
ARM object code directly to an object file or even directly into memory.
> 
> What about the performance of the generated code compared with
> ST->C->Assembler generation by Squeak C codegenerator followed by gcc?
In the only case I can remember, the full VM generated by the
PurdyTranslator was something like 10% faster than the CCodeGenerator->C
VM. 
> 
> Last but not least: What about license issues?
Craig & I were given permission to put it all under the Squeak license.


tim






More information about the Squeak-dev mailing list