The biggest change is in the way lexical scopes are handled. Lexical scopes basically consist of two chunks of data: static information like type details and so on, and dynamic data, which basically amounts to the current value of variables and such.
Previously, whenever a new scope was entered, the VM would create a complete copy of the scope, including the static data. This meant spending a lot of time copying trees and arrays around, which obviously isn't optimal. So I refactored the scope system so that dynamic information is now stored in a separate class entirely. The static data is no longer cloned, which gives us a healthy speed boost and a bit of memory savings as well.
I've also hacked in a variety of other fixes and improvements, many of which came as fallout from the scope handling changes. I'm not convinced that the code is truly optimal yet (there's some wizardry that could be done to improve cache coherency and some other stuff) but it is pretty damned fast.
For comparison, in my last post, I mentioned that stress testing reveals that 60,000 messages can be sent per second between threads. That number is now hovering around 102,000. Most of that was accomplished by minimizing the scope performance overhead, because each message handled represents one scope entry and one scope exit.
I've played around with some special allocators for data that tends to have heavy churn and short lifetimes, such as RValue wrappers; so far it looks like there won't be any benefit to writing special allocators in general. There's a few places where they work well (including a spot in the messaging system) but overall, due to thread safety issues, it's just simpler and easier to let the standard library do its thing directly.
On another front, I discovered a brilliant little Schroedinbug in the external DLL call system. For some mystical reason, the compiler produced the exact right code to keep things from blowing up. In working on something totally unrelated, I disrupted the careful balance, and the compiler started emitting code differently.
The new code stomped randomly all over the stack, and produced some truly horrific crashes that were inconsistent, difficult to reproduce, and often manifested simply as programs giving the wrong results instead of outright exploding.
As it turns out, the issue was related to the way I pushed parameters onto the stack when invoking external DLL functions. Pushing parameters shifted the stack pointer; when the code later tried to use the stack pointer to access a local variable, all hell broke loose, doing wonderfully random and unpredictable stuff to the running app.
I managed to get around the problem by rewriting the DLL call invocation routine mostly in assembler, which seems to be holding pretty solid.
So although some major headway has been made on the VM, I failed to check off any major items on the TODO list. In fact, I've just run a search of the codebase for TODO items, and the revised list comes up something like this:
- Integrate new features with the bytecode system
- Respect 16-bit wide parameters when invoking DLL calls
- Double check operator precedences
- Finish support for user-defined infix operators
- Fix a bug where types are ignored during infix computations
- Retest all infix assignment variants for correctness
- Fix bug with prefix unary operators applied to non-literal values
- Pop the parser stack correctly; many sections of the parser are guilty of leaving their junk hanging around, making error reporting messy
- Implement support for named lists
- Type safety when passing list parameters
- Fix(?) boost static_assert weirdness
- Implement allocators properly
- Remove allocator hacks
- Make x++ and related operations work correctly
- Vomit if we allocate too many messages
- Better handling of name collision detection
- Reorganize some files
- Consider trading off code duplication for speed in a few cases
- Make it legal to access global constants from a task
- Fix some bugs in nested response map support
- Improve syntax for nested structure initialization
- Type aliases (aka. typedefs)
- Change task IDs to string variables for easier metaprogramming
- Perform complete code review for exception safety, documentation, code cleanliness, error handling robustness, and elimination of hardcoded strings/magic numbers
So yeah. R7 is moving along fast, and will still take a billion years to actually arrive. We will literally evolve the ability to implement entire programming languages with our toes before I finish this release. But the release will occur.
Eventually.
