Runtime Documentation

Differences with MLton as of 10/5/17

• Main compiler hasn't diverged much
  • We have:
    • read/write barriers
    • limit check
    • enabled polymorphic FFI
  • They have:
    • Quality-of-life updates (better error messages?)
• Runtime is Linux-only
  • Only support C backend (can't do LLVM, native, ...)
• Interrupts (for multi threading)

Per-file Documentation

include/c-main.h

Why is main a macro?

• In generated C code, this file is included.
• This allows targeting different architectures? (Swapping out different implementations of main)

main

• pthread_key allows pthread local storage
• GC_state is a catch-all for all necessary processor-local storage
  • statistics is per-processor
  • global heap is shared
  • Duplicate chooses whether or not to copy (for local vs shared)

MLton_threadFunc

• s->amOriginal
  • For world saving; is probably broken at the moment.
• conts are like basic blocks
• trampoline executes conts (this is the whole program from P0's perspective)
• Other processors wait until runtime call which starts them up

runtime/gc/garbage-collection.c

GC_collect

• called every time we're sure if we have enough memory (called by any processor)
• force = true when, e.g., MLton.GC.collect()
• getStackCurrent
  • used field is a length (in bytes) of how big the stack is
  • used is not updated during normal stack manipulation, so it needs to be updated before GC stuff
• beginAtomic
  • is not atomic across pthreads, but atomic within my pthread
  • guards against signal interrupts!
  • signal handlers are also ML code

Misc

• There is a runtime call, MLton exit, which is called when the SML program terminates
  • Called regardless of nature of exit, e.g., unhandled exceptions (but not segfaults)
• What is parallel_benchmark directory? We never touched this directory.