mirror of
https://github.com/grin-compiler/grin.git
synced 2024-09-11 15:37:54 +03:00
22 lines
650 B
Plaintext
22 lines
650 B
Plaintext
The whole program AST could contain few million nodes. i.e. lambdacube-quake3 has ~1.1M nodes
|
|
|
|
This poses serious challenges for performace requirements.
|
|
|
|
Q:
|
|
- how far can we get with the current system design? (CPU, ShortText names, pointer tree)
|
|
problems:
|
|
- memory footprint
|
|
- traversal speed
|
|
- data-flow analysis speed
|
|
possible improvements:
|
|
- multithreading
|
|
- int names + symbol table
|
|
- c interpreter for data-flow machine
|
|
- llvm jit for data-flow machine
|
|
|
|
Possible system designs:
|
|
- data-flow analysis on GPU
|
|
- packed AST representation
|
|
- optimization passes on CPU with SPMD
|
|
- optimization passes on GPU
|