DIGITALIS

Two-Gear JIT Compiler

Translation happens in two gears. Both emit native x86_64 machine code, and both cache what they produce, so a region is translated once and then simply run.

• First gear — the Lite Translator. A single pass over a cold region, covering ~98% of the instruction set with no optimisation. Its job is to get working native code as quickly as possible — and to keep regions whole: the register allocator holds a fixed six-register scratch reserve so that running out of registers spills through memory instead of splitting the region, an invariant that eliminated 70,000+ measured region splits at function prologues across a 14-app sweep.
• Second gear — the Heavy Optimizer. Re-compiles a region once it has proved hot, with global register allocation and loop optimisation across the whole region. Neutral-or-faster than first gear in general, up to ~2x on register-pressure-heavy loops.
• Checked against itself. 3,600+ host translation tests, plus differential fuzzers that run the same code through the JIT and the interpreter and compare the entire CPU state afterwards.

Interpreter Fallback

Per-instruction execution for syscalls, complex SIMD, and anything the JIT can't handle. Ensures complete ARM64 coverage.

Proxy Libraries

A guest ARM64 library cannot call the host's x86_64 one directly. 21 proxy libraries stand on that boundary — Vulkan, libc/libm, EGL/GLES, AAudio, camera, NDK binder, NNAPI, JNI helpers — translating each call and forwarding it to the native implementation.

• Graphics reach the real GPU. Vulkan goes through GFXStream's VkDecoder; OpenGL ES is driven by ANGLE, which implements it on top of Vulkan in-guest — giving full OpenGL ES 3.2 and hardware MSAA on the host GPU.
• Awkward symbols are covered by hand. Anything the auto-generated marshallers can't handle gets a custom trampoline written in-surface — JNIEnv translation, host→guest callbacks, and the ~80 ANGLE extension procs a browser engine gates on but eglGetProcAddress would otherwise return NULL for.
• The boundary is where the subtle bugs live. A zero stride from ANativeWindow_lock made a video app render solid green; a guard-page read inside the allocator wrappers looked for weeks like host memory exhaustion. Both fixed here — and the stride now comes from asking the host allocator rather than assuming the format's rule, because row padding is gralloc's choice, not the pixel format's.
• Probed systematically, not just when apps break. Eight dedicated samples now stress the riskiest marshalling paths — host-invoked guest callbacks, buffer handles crossing unix sockets, fd passing, JNIEnv round-trips — and all of them pass; the only failures they ever found reproduced identically on native x86_64 builds, i.e. they were Android platform behaviour, not translation.
• No reachable call aborts, audio included. The OpenSL ES and OpenMAX AL proxies wrap host interface vtables method by method, and the methods upstream could not marshal used to kill the process on first call. A named-override seam now supplies working marshalling for every one an app can reach — the streaming-decode buffer-queue callback fires from host code back into translated guest code — closing the last known app-reachable aborts across all 21 proxies.

150 Sample Apps

Every sample is an ARM64-only APK that runs end to end on the x86_64 emulator: ported NDK samples, third-party native libraries, whole UI engines, and one probe per ARM64 instruction extension.

• Real libraries, linked the way an app links them. OpenCV, TensorFlow Lite, PyTorch, ONNX Runtime, FFmpeg, WebRTC, Qt 6, React Native, Realm, SQLCipher and dozens more — not reimplementations written to be easy to translate.
• A wrong answer fails, loudly. The ML, audio, crypto and binder samples check their output bit-exact against a hardcoded golden, so a miscompiled NEON kernel aborts instead of quietly rendering something slightly wrong.
• One sample per extension. NEON, FP16, BF16, dot-product, I8MM, complex arithmetic, JSCVT, PAC, BTI, the LSE and LRCPC atomics, CRC32 and the crypto instructions each get a probe that runs hot enough to reach the optimising JIT tier.
• One probe per proxied NDK surface. Eight samples exercise the full public API of a proxy library each — AHardwareBuffer, AImageDecoder, the media data-source/muxer pair, the ADPF hint/thermal/choreographer trio, ASharedMemory, system fonts, OpenMAX AL, OpenSL ES — with golden self-checks on every call, so a marshalling bug fails a specific assertion instead of corrupting an app somewhere downstream.

The full catalogue is grouped by subsystem further down the page.