March 2026 perf improvements

[struct]

Performance fixes from Claude!

iso_alloc_zone_t field reordering — include/iso_alloc_ds.h
The biggest win. is_full was previously at offset ~2,119 bytes (cache line 33), buried after the 2,040-byte free_bit_slots[255] array. Every call to is_zone_usable() — the first check in the hot allocation path — would miss the cache loading that field.

The new layout puts all hot fields (user_pages_start, bitmap_start, next_free_bit_slot, canary_secret, pointer_mask, max_bitmap_idx, chunk_size, free_bit_slots_usable, free_bit_slots_index, is_full, internal) in the first 64 bytes (one cache line). The large free_bit_slots[255] array, which is only accessed during free-list refills, moves to the end.

__builtin_ctzll in iso_scan_zone_free_slot_slow — src/iso_alloc.c
Replaced all inner for(j = 0; j < 64; j += 2) loops with:

uint64_t free_mask = ~(uint64_t)bts & USED_BIT_VECTOR;
if (free_mask) return (offset + __builtin_ctzll(free_mask));
USED_BIT_VECTOR = 0x5555... selects even-position bits (one per chunk). Inverting + ANDing gives a mask of free slots. CTZ finds the first in one instruction instead of 32 iterations. Applied to all three paths: NEON, __int128 (split into two 64-bit halves), and standard.

__builtin_ctzll in fill_free_bit_slots — src/iso_alloc.c
Same technique for populating the free-list cache, replacing the 32-iteration inner loop in the partial-word case with a free_mask &= free_mask - 1 iteration (classic "iterate over set bits" idiom).

Zone cache scan direction — src/iso_alloc.c
Changed the thread zone-cache scan from oldest-first (0 → count-1) to newest-first (count-1 → 0). The cache is populated LIFO, so the most recently used zone — most likely to still have free slots — is found on the first iteration instead of the last.

[jvoisin]

Are there any benchmarks to validate those claims?