Important: this changes the semantics of the write barrier!
Carefully read the big comment block in lj_obj.h
This helps HREFK key slot specialization and allows safely hoisting
HREF/HREFK across GC steps, too (fix for a barely reproducible bug).
Dead keys are only removed during a table resize (as before).
Need to sync GC objects to stack only during atomic GC phase.
Need to setup a proper frame structure only for calling finalizers.
Force an exit to the interpreter and let it handle the uncommon cases.
Finally solves the "NYI: gcstep sync with frames" issue.
Record __call resolving and specialization for CALL* bytecodes.
Record argument adjustment and fast functions for FUNC* bytecodes.
Avoids all pending/immediate decisions for chained fast functions.
Cleaner semantics for pcall(), xpcall() and __tostring metamethod.
Prerequisite to drop the shadow frame link stack again.
Drop call gates. Use function headers, dispatched like bytecodes.
Emit BC_FUNCF/BC_FUNCV bytecode at PC 0 for all Lua functions.
C functions and ASM fast functions get extra bytecodes.
Modify internal calling convention: new base in BASE (formerly in RA).
Can now use better C function wrapper semantics (dynamic on/off).
Prerequisite for call hooks with zero-overhead if disabled.
Prerequisite for compiling recursive calls.
Prerequisite for efficient 32/64 bit prototype guards.
Typically reduces storage overhead for snapshot maps by 60%.
The extensible format is a prerequisite for the next redesign steps:
Eliminate IR_FRAME and implement return-to-lower-frame.
All upvalue objects hold a disambiguation hash value now.
It's built from the parent prototype and the slot number.
Different hash values imply the upvalues cannot alias.
Same hash values don't imply anything (collision or different closures).
Upvalue disambiguation makes use of a reduced hash due to IR contraints.
Compile math.sinh(), math.cosh(), math.tanh() and math.random().
Compile various io.*() functions.
Drive the GC forward on string allocations in the parser.
Improve KNUM fuse vs. load heuristics.
Add abstract C call handling to IR.
