NeonJS

mooner:~/resume/projects/neonjs $ tree -L 1

mooner:~/resume/projects/neonjs $ cat docs/architecture.md

mooner:~/resume/projects/neonjs $ git log --grep=fix

1. 01

   The runtime aborts on Android 8.0 and 8.1

   Problem

   The API 26 emulator CI died on a JIT thread with Check failed: self == thread_running_gc_. ART 8.x records the only implementation of an interface method and reads that record again even after the class loader behind it is gone, and the probe class loaded first was that only implementation.

   Approach

   Two implementations of CompiledCode are now linked in the engine’s loader up front, so ART stops tracking the method. On those versions Java.extend adapters stay loaded, and the device script prints the native backtrace from crash logs.

   Result

   The next Android CI run passed. Today the API 26 and 34 emulators pass all 7 device checks and a 15,789-test Test262 subset.

2. 02

   More than 750,000 JIT classes

   Problem

   JIT class names hashed only the instructions and register count, so in Test262 78 names stood for genuinely different code. That ruled out putting several classes in one dex, and the same code got a new class for every block.

   Approach

   Names became a SHA-256 of everything the generator reads, so equal names mean equal class files. Generated classes hold no state, so blocks with the same input share one instance from a weakly held cache.

   Result

   As measured in the commit, Test262 in compiled mode went from 756,204 generated classes to 75,915, and from about 150 s to 35 s. 51,915 names match 51,915 distinct class files.

3. 03

   The fixed cost of dex conversion adds up per function

   Problem

   On Android, JIT classes must be converted to dex, and every dx or D8 call has a large fixed cost. Converting each function on its own pays that cost once per function.

   Approach

   Compilation moved to background daemon threads that take up to 32 blocks at a time and turn each batch into one dex file and one class loader. Exactly one thread compiles a block, a rejected batch is defined class by class, and worker time is not charged to a script’s limits.

   Result

   As measured in the commit, compiling 5,000 functions at once on a desktop JVM went from 1,331 ms to 671 ms with dx and from 7,150 ms to 2,400 ms with D8. Through dx, all 48,747 Test262 tests still pass.

4. 04

   Compiled code boxes a number at every step

   Problem

   Generated code held every value as an Object, so each arithmetic step allocated a new Double.

   Approach

   A forward type analysis gives registers and stack slots a kind (double, int32 or anything else), and registers that only hold numbers become JVM locals. Ints are used only for values that are int32 by definition, such as bitwise results, and arithmetic stays in doubles, so overflow and -0 behave as before.

   Result

   As measured in the commits, on an ART emulator an empty loop went from 9.22 to 0.70 ns per iteration and xorshift32 from 25.3 to 3.5 ns. 60,000 random numeric functions compared against the interpreter showed no mismatch, and Test262 passes with unboxing on and off.

5. 05

   HotSpot and Android call different methods

   Problem

   Among overloads of equal conversion cost, the one reflection listed first won, and HotSpot and ART list methods in different orders: sb.append(null) called a different method depending on the VM. Long IDs beyond 2^53 also changed their last digits on the way to a JS number.

   Approach

   Ties now go to the most specific types, as in Java, then to a fixed order, and a signature can name the overload wanted. Longs beyond 2^53 − 1 cross over as BigInts.

   Result

   Both are pinned by the interop contract tests, and the Android device check covers the long boundaries.

6. 06

   The sandbox does not do what its docs say

   Problem

   The security docs said waiting for a busy context was bounded by the time limit, but the wait had no bound and two threads could deadlock. The engine’s own API classes were missing from the deny list, and new URL('a:#' + 'x'.repeat(2e6)) ran a quadratic loop that no time limit could stop.

   Approach

   Lock waits are bounded by the time limit and end in a NeonTimeoutException; every engine package and API class is on the deny list; the URL parser is linear and checks for interrupts and limits every 4,096 steps.

   Result

   All three are pinned by tests. A new class in the engine packages must be denied or declared harmless in the test.