Inline caches
Opened this issue · 2 comments
We need to collect type information in inline caches.
They should be stored in a Code space of heap.js. The ICs should be used only for a baseline compiler to gather the type info for optimising compiler.
I think we should start with simple ICs without type feedback collection. Since this is quite huge task it'll be better to move step by step. At first I suggest to start with unary/binary op stubs then implement property lookup/store (it's still quite unclear for me how we will optimize here - v8 hidden classes approach or something else?).
The proposed algorithm for the first iteration (monomorphic binary/unary op ICs for Numbers without type feedback collection) is quite simple:
- At first we have just generic dispatching stub we go through looooong sequence of conditions and try to seek appropriate specialized stub (uninitialized state).
- Install specialized stub then call it (monomorphic state).
- In specialized stub we perform simple assertions on operand type tags. If assertion fails we fallback to generic stub. If it passes assertions then we perform type-specific operation and tag result.
So, TODOs for first iteration for me looks like this:
- Implement tagging for all primitives.
- Figure out how to install/bailout stubs with minimal perfomance penalty. Using call with all the call-convention juggling doesn't look like a good option. However, otherwise we will need to recompile whole scope CFG with installed stub and re-run linearscan which is definitely sucks too. Any ideas?
- Figure out if we can create somekind of IC generator instead of hand-written approach. As I see it: we can create hand-written stubs for same type operands (e.g. Int32 and Int32) with type-conversion place-holders. Then impement type conversion stubs. So, then we hit (Int32, String) we need generate (String, String) stub and inject (Int32->String) conversion stub for first argument.
- Implement same-type stubs for Int32 and/or Smi
- Implement same-type stubs for Double
- Implement conversion stubs for Numbers
Update:
- Benchmarks, benchmarks, benchmarks!
I think this should be enough to figure out techniques and approaches which we will use later. Then we can plan second iteration.