For the varargs functions that used the Type field to validate their parameters, now a hidden additional argument is passed which contains a byte array with the type info for the current call's arguments. Since this is static per call location it can be better prepared once when the code is being compiled instead of being put in a runtime created array for each invocation. Everything else uses the per-function instance of the same data.
The only thing that still needed the type field with a VMValue is the defaults array, so this uses a different struct type now to store its data.
The amount of support code for this minor optimization was quite large and this stood in the way of streamlining the VM's calling convention, so it was preferable to remove it before moving on.
With a proper count value available this can be done properly. The only relevant targets are the jumps immediately succeeding the IJMP instructions, nothing else.
The instruction one free instruction byte so it's now using that to extend its argument's register range to 65535.
For param this is needed because it passes strings by reference and creating an implicit temporary copy for string constants does not work here.
Combining these two groups of data has been the cause of many hard to detect errors because it allowed liberal casting between types that are used for completely different things.
This is an incredibly costly way to do a debug check as it infests the entire VM design from top to bottom. These tags are basically useless for anything else but validating object pointers being passed to native functions (i.e. mismatches between definition and declaration) and that simply does not justify a feature that costs execution time in non-debug builds and added memory overhead everywhere.
Note that this commit does not remove the tags, it only discontinues their use.
This makes VMValue a real POD type with no hacky overloads and eliminates a lot of destructor code in all places that call a VM function. Due to the way this had to be handled, none of these destructors could be skipped because any value could have been a string.
This required some minor changes in functions that passed a temporary FString into the VM to ensure that the temporary object lives long enough to be handled. The code generator had already been changed to deal with this in a previous commit.
This is easily offset by the code savings and reduced maintenance needs elsewhere.
Both files can now be included independently without causing problems.
This also required moving some inline functions into separate files and splitting off the GC definitions from dobject.h to ensure that r_defs does not need to pull in any part of the object hierarchy.
- block creation of actors with the 'new' instruction. Unlike the above these cannot be made abstract because without ConstructNative they cannot be serialized.
It now uses a dedicated opcode instead of piggybacking on OP_CALL and it passes data that is closer to the VM. Symbols should be avoided at this level.
It also will skip the scope instruction if the code generator detects that both calling function and the self pointer type have the same scope, this assumes that subclasses cannot flip between UI and Play.
They are not needed for OP_NEW_K which can evaluate the class relations at compile time and for OP_NEW the calling function can also be checked at compile time, passing only the scope value itself.
