[alloc+atomic] Optimize GhostTreeBStackAllocator

[williamwutq]

Description: Bring several optimization to GhostTreeBStackAllocator and its corresponding C implementation by child pointer and height caching, rotation threading, reducing dynamic heap allocation on hot path, and other techniques. In addition, allow C fuzzing to skip fsync, mirroring similar Rust optimization.

Important Feature: No
Type: Allocator - Optimization
Magic Number: ALGT
Bulk: No
Tests: Included
Feature Flags: alloc + set
Breaking change: No
New Types: None
Rust Only: No
Fuzz: Yes
Safety Review: Needed: Crash Safety, Invariants, Thread Safety

[williamwutq]

Perf data on my machine (mixed/uniform)

threads	deep-tree (iters=90) master→alloc	ratio	median ratio
1t	37.9 → 26.3 ms	0.69	0.73
2t	44.0 → 31.1 ms	0.71	0.72
4t	38.6 → 25.9 ms (iters=72)*	0.67	0.67
16t	45.2 → 31.0 ms	0.68	0.75

[williamwutq]

Per-op latency (ms), 10 samples/case, iters 9→90 on APFS SSD (my machine)

threads	branch	min	P20	P40	P60	P80	max	deep-ratio a/m
1t	master	22.56	28.64	33.99	43.10	53.79	64.10	—
1t	alloc	18.55	20.77	24.42	30.33	36.89	42.75	0.69
2t	master	24.39	31.80	44.76	47.50	51.64	68.66	—
2t	alloc	21.00	26.44	28.03	33.19	38.18	46.98	0.68
4t	master	16.41	18.62	27.70	31.44	35.50	45.74	—
4t	alloc	17.55	20.25	21.65	29.01	33.61	38.79	0.78*
16t	master	19.65	27.81	32.95	45.64	51.39	67.86	—
16t	alloc	13.61	17.77	21.58	26.47	35.61	39.48	0.75

*4t deep-ratio is measured at iters=54 (criterion picked a different iter step, so the deepest common tree is shallower); its paired-median over common iters is 0.68, in line with the others.

This shows average ~27.5% lower per-op latency, a great improvement.

[Copilot]

Pull request overview

This PR optimizes the GhostTreeBstackAllocator (Rust + C) hot paths by avoiding per-operation heap allocations and reducing I/O during AVL updates via cached child heights, while also adding a C-only test/fuzz escape hatch to skip durable sync for faster fuzzing.

Changes:

• Add denormalized cached child-height fields to AVL nodes and thread known heights through insert/remove/rotations to avoid extra reads.
• Replace Vec-allocated path buffers with fixed-size stack arrays on AVL operations (insert, remove-min, best-fit remove).
• Add BSTACK_TEST_NO_DURABLE_SYNC to make plat_durable_sync a no-op in C test/fuzz builds; remove the now-implemented optimization plan section from PLANNED.md.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
src/alloc/ghost_tree.rs	Adds child-height caching + stack-based path tracking to reduce allocator critical-section work and I/O.
c/bstack_alloc.c	Mirrors the Rust AVL optimizations (cached heights, threaded rotations, stack paths) and bumps ALGT magic.
c/bstack.c	Adds compile-time test/fuzz option to skip durable sync for faster fuzzing.
PLANNED.md	Removes the planned item that is now implemented by this PR.

Suppressed comments (1)

c/bstack_alloc.c:2376

• nbf and nh are written by algt_read_node but never read afterwards in this branch either, which can trigger -Wunused-but-set-variable warnings. Consider explicitly marking them as used (or avoid reading them if not needed).

        uint64_t size, left, right, right_sz, right_l, right_r;
        int8_t nbf, right_bf; uint8_t nh, right_h;
        if (algt_read_node(bs, node, &size, &nbf, &nh, &left, &right) != 0) return -1;
        if (algt_read_node(bs, right, &right_sz, &right_bf, &right_h, &right_l, &right_r) != 0)

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.