Skip to content

VM_PERFORMANCE_UPDATE_APRIL_2026.md

Latest commit

 

History

History
61 lines (48 loc) · 2.97 KB

VM_PERFORMANCE_UPDATE_APRIL_2026.md

File metadata and controls

61 lines (48 loc) · 2.97 KB
Raw

ProXPL VM Performance Milestone (April 2026)

This document outlines the major architectural optimizations implemented to accelerate the ProXPL Virtual Machine. These changes represent a transition from a baseline interpreter to a high-performance threaded runtime.

🚀 Key Optimizations

1. Direct Threaded Dispatch Fix

Fixed a critical architectural bug in the experimental threaded dispatchers where the jump table was being rebuilt every iteration.

  • Files: src/vm/vm_dispatch.c, src/vm/vm_core_opt.c
  • Improvement: Jump table is now initialized once before the hot loop.
  • Impact: Enables true computed-goto performance (~1.5x-2x speedup on GCC/Clang).
  • Safety: Added do_INVALID trap handlers to all 256 opcode slots to prevent undefined behavior on unrecognized bytes.

2. Register Caching (ip & stackTop)

The primary execution loop now caches the most volatile pointers in local CPU registers.

  • File: src/runtime/vm.c
  • Details:
    • ip (Instruction Pointer) and stackTop are now local register variables.
    • Reduced memory traffic by avoiding constant structure dereferencing.
    • Implemented STORE_FRAME() and LOAD_FRAME() synchronization primitives for C-call safety.

3. Global Inline Caching (GIC)

Implemented a single-level inline cache for global variable lookups to bypass hash table overhead.

  • Infrastructure: Added cache field to ObjFunction and GICEntry structure in vm.h.
  • Logic:
    • On first access, the variable's entry address and table pointer are cached.
    • Subsequent accesses verify the table pointer (to handle rehashes) and perform a direct memory read.
  • Impact: Turns $O(n)$ hash probes into $O(1)$ direct access.

4. Hash Table Arithmetic Optimization

Optimized the core mapping logic in the Table implementation.

  • File: src/runtime/table.c
  • Optimization: Replaced expensive modulo (%) with bitwise AND (&) for indexing.
  • Prerequisite: Ensured all table capacities are powers of two (already guaranteed by growth logic).

📊 Performance Matrix (Estimated)

Optimization Target Metric Expected Gain
Threaded Dispatch Instruction Overhead 30% - 50%
Register Caching Memory Traffic 15% - 20%
Inline Caching Global Variable Speed 200% - 400%
Bitwise Indexing Table Lookup Latency 5% - 10%

🛠️ Internal Changes for Developers

  • ObjFunction: Now includes a void* cache which is GC-managed (freed in gc.c).
  • Table: New tableGetEntry() function provides direct Entry* access for caching systems.
  • VM Loop: All opcode handlers in run() must now use local ip and stackTop and call STORE_FRAME() before any operation that might trigger a GC or use the VM's global state.

Status: Milestone 1 Complete