For now, this GC is non-generational, and much slower than the old
version. It tracks objects by a fixed object ID rather than changeable
memory address. Small object (eight bytes or less) are stored directly
in the array, indexed by object ID, while larger object are allocated
with malloc() (for now) and stored in the array as a pointer. Object
IDs are stored as 32-bit integers, even on 64-bit platforms.
Advantages:
- Simpler design
- Requires less memory on 64-bit platforms
- Object IDs don't change when running the GC
- No need to store a random "hash" value in vectors/strings/structs
- Can hash pairs by identity, not just value
- Can move objects individually, without fixing up all references
- Can determine object type from value, without another memory access
Disadvantages:
- Lower initial performance (non-generational, relies on malloc())
- 32-bit values place a (high) limit on total number of objects
- Must explicitly free unreachable object IDs after GC
Each transient identifies the value of the corresponding previous bytecode.
This change (a) frees up many bytecodes formerly used by the conditional
expression (if c t f); (b) regularizes the bytecode by always placing opcodes
before operands; and (c) causes the bytecode to conform to the Single Static
Assignment (SSA) form preferred by e.g. LLVM.
Includes updates to the hand-assembled files (*.rla) and the bytecode compiler.
Eliminate use of #="undefined" as an explicit initializer for boxes.
Do not allow #@ ("freeze") to be applied to references, for sanity's sake.
Inside compiler, builtins are now represented by (#%builtin "name") form.
Plain symbols are promoted to builtins; quoted symbols become structures.
Fix choice of continuation when function ends in (call/cc).
Change empty kw-args from '() to #%nil in (simplify-apply).
Improve error message when no match is found for a primitive form.
Add special variables for keyword arguments & values.
Add support for keyword arguments to (simplify-apply).
Implement full string parsing in src/reader.rls.
TODO: Support keywords in (simplify-lambda) and writer.
Fix assertion in gc_object_left_behind() to work in Gen-0 collection.
Add (weak-box?) and (make-weak-box) primitives. (unbox) now works for both.
Add option to just simplify the input (stops before reduce-function).
Default to writing .rla without indentation, comments, or newlines.
Add support for (fix=), (list), (and), (or), (cond), (when), and (unless).
Fix a mapper bug which could assign the same frame var to separate variables.
Update make-struct primitive for new structure type layout.
Change primitives to use #% as prefix instead of just %.
Add primitive operations for comparing byte-strings.
* Remap all instance variables in one pass, to avoid an insidious bug [1].
* Fix another bug by merging promote-shared-variables and narrow-binds [2].
* Don't assume different variable means different value in propogate-set!.
* Add support for (apply), (value-list) and (call-with-values) forms.
* Add support for including files (as if directly substituted) in reader.
* Use scheme/match library to simplify form pattern-matching.
* Refactor (map-form) and (search-form) using a more basic (traverse-form),
which just recurses over the form and returns void by default, and a
new utility function (curry-keywords) to provide default keyword arguments.
[1] Was renaming e.g. %f0 to %i0, then %i0 to %i1, which eliminates the
distinction between %f0 and %i0. Solution is to construct a map from old
names to new names, then traverse the form and change every old variable
to its new equivalent in the map exactly once.
[2] Some variables were not being promoted to boxes, as promotions only occur
at the top-level, when constructing each lambda, and narrow-binds could push
the unpromoted variables down into a subordinate lambda form first. Solution
was to promote variables immediately after narrowing bindings, including the
recursive calls which exist after pushing variables into nested scopes.
Created one module one per pass, plus utilities functions, primitives, and output.
Changed extension to ".scm" for compatibility with hg syntax highlighting backend.
Jesse D. McDonald