Repository navigation
Skyrim Androidborn 0.2.3
0.2.3 — 2026-11-08
Performance pass for heavy world scenes.
Performance
- [x87 top of stack kept in a local register — about 20% less CPU time per frame in CPU-limited scenes]
The common fld/fstp/fmul/fadd path no longer reloads the FPU stack pointer for every instruction. - [Internal symbols hidden and PLT calls removed — less overhead on every recompiled call and a faster start]
Generated functions no longer go through the dynamic linker. - [Interlocked operations without redundant fences — fewer stalls in busy scenes]
The atomic read-modify-write already provides the ordering x86 needs. - [Watchdog diagnostics kept out of the hot path — lower CPU use while playing]
The per-import record only keeps the cheap last-call fields during normal play. - [Guest memory asks for huge pages — fewer TLB misses where the kernel supports them]
The 4 GiB guest window uses 2 MiB pages when the kernel allows them. - [ARMv8.2/LSE build — faster locks and reference counts on modern phones]
Interlocked operations compile to single instructions instead of outline helper calls.
Notes
This build targets ARMv8.2/LSE (Cortex-A75/A55 and newer). The Retroid Pocket
G2 and Xiaomi 15T are supported; older ARMv8.0 devices should stay on 0.2.2.
At 1280x720 low in Whiterun, where the CPU limits the frame rate, 0.2.2 ran
at 39.6 fps and 0.2.3 at 49.6 fps; CPU time per frame fell about 20%. At
1920x1080 with 2x MSAA and 4x anisotropic filtering the GPU is the limit:
both versions stay around 38 fps at the same GPU clock and similar
temperatures. The gain is CPU headroom for lower resolutions and settings,
not extra frames when the GPU is already saturated.