From 29fb6aa1eb0f9551d1cbadd6cc8a9ae50b4480d3 Mon Sep 17 00:00:00 2001
From: Konstantin Nazarov <mail@knazarov.com>
Date: Tue, 15 Aug 2023 21:41:39 +0100
Subject: [PATCH] Add a post about garbage collection

---
 .../note.md                                   | 69 +++++++++++++++++++
 1 file changed, 69 insertions(+)
 create mode 100644 content/posts/vm_progress_update_simple_garbage_collection/note.md

diff --git a/content/posts/vm_progress_update_simple_garbage_collection/note.md b/content/posts/vm_progress_update_simple_garbage_collection/note.md
new file mode 100644
index 0000000..62d034e
--- /dev/null
+++ b/content/posts/vm_progress_update_simple_garbage_collection/note.md
@@ -0,0 +1,69 @@
+X-Date: 2023-08-15T21:30:00Z
+X-Note-Id: 81666e7b-ed56-400a-b08d-5edad9f966f5
+Subject: VM progress update: simple garbage collection
+X-Slug: vm_progress_update_simple_garbage_collection
+
+Today I've reached another milestone with the virtual machine.
+It can now allocate memory and perform garbage collection.
+
+It is interesting that I don't yet have a high-level language implemented,
+but the garbage collection is still useful. In bare-metal assembly, you have
+to track memory allocations and deallocations yourself, since the memory is unstructured,
+and a potential garbage collector won't be able to differentiate between pointers and
+integers. In my virtual machine (and its assembly language), all memory is tagged and
+structured, and thus if you have a pointer to the virtual machine data structure, the
+garbage collector is able to walk it. So even the assembly/bytecode will soon be able to
+work as if memory is not a concern.
+
+For the implementation of garbage collection, I've chosen a modified
+[Cheney's algorithm](https://en.wikipedia.org/wiki/Cheney%27s_algorithm).
+This algorithm contains two distinct "heaps", one of which is currently in use for allocation.
+If it runs out, the algorithm would walk all objects that are reachable from the registers and
+the stack of the virtual machine, and move them to the other heap. The previous heap would then
+contain only dead objects, and can be grown/shrunk for the next cycle if needed.
+
+For any practical purposes, Cheney's algorithm would be quite slow, and many real programming
+languages contain multiple heaps (python has 3) and vary the size of the heaps to reclaim
+short-lived objects faster. However, I predict that the naive algorithm would get me pretty far along
+the path to the implementation of the actual language, so I don't care that much about it now.
+
+What follows below is a piece of code that demonstrates the C interface for interacting with
+VM memory allocation and GC capabilities. It is taken almost verbatim from the unittest and
+annotated.
+
+```
+# Allocate all structures required by the virtual machine
+vm_t* vm = new_vm();
+
+# Allocate an array of 10 elements
+tagged_value_t arr = vm_mk_array(vm, 10);
+array_set(arr, 0, mk_i64(42));
+
+# Save pointer to the array in a register so it survives garbage collection
+vm->registers[R1] = arr;
+
+assert(arena_generation(&vm->arena, arr) == 0,
+       "Array should be in first generation\n");
+
+# Perform garbage collection
+vm_gc(vm);
+
+# The array has moved, so retreive a pointer to it from the register
+arr = vm->registers[R1];
+
+# Check that the array contains correct value at 0-th position
+tagged_value_t val = array_get(arr, 0);
+assert(op_eq(val, mk_i64(42)), "0-th element of an array == %ld\n",
+       val.value.i64);
+
+# Make sure that the array has really moved between generations
+assert(arena_generation(&vm->arena, arr) == 1,
+       "Array should be in second generation\n");
+```
+
+The API is in place, so the next step would be to add a few instructions to the bytecode
+that can call memory allocations, and write an inefficient merge sort (the one that would
+copy sub-ranges).
+
+As usual, the code can be found [here](https://git.sr.ht/~knazarov/lisp.experimental)
+(note that it's experimental, so don't expect any clarity).