Abstract:

This article announces the release of Julia version 1.12, detailing a range of significant new features and improvements aimed at enhancing performance, developer experience, and tooling. Key highlights include the experimental --trim feature for creating smaller, faster-compiling binaries by eliminating unreachable code, and the ability to redefine constants and structs within a live session, which dramatically improves iterative development workflows.

The release introduces new multi-threading capabilities, such as a default interactive thread for a more responsive REPL and respect for CPU affinity settings in containerized environments. Performance analysis is made easier with new tracing flags and macros for inspecting compilation and dispatch. The package manager, Pkg, gains major features like "Workspaces" for monorepo management and "Apps" for creating standalone command-line tools. Additional improvements include build optimizations with BOLT, enhanced atomic operations, more reproducible testing via RNG state reporting, and foundational changes to LLVM IR generation for better low-level interoperability.

Julia 1.12: Key Features and Enhancements

• --trim Feature: An experimental flag for JuliaC.jl that enables static code trimming. This removes unreachable code to significantly reduce binary size and compilation time, ideal for creating lightweight applications.
• Redefinition of Structs: A major enhancement for interactive development, allowing constants and structs to be redefined without restarting the Julia session. This feature integrates with Revise.jl to make prototyping much smoother.
• New Compilation Tracing Tools:
  • --trace-compile-timing flag shows how long each method takes to compile.
  • @trace_compile and @trace_dispatch macros allow for ad-hoc, targeted tracing of compilation and dynamic dispatches for a specific expression, aiding in performance analysis.
• Multi-Threading Improvements:
  • Default Interactive Thread: Julia now starts with one default thread and one interactive thread (-t1,1), ensuring the REPL remains responsive even when user code is running.
  • CPU Affinity: Julia and its BLAS threads now respect CPU affinity settings (e.g., from Docker, cgroups, taskset), preventing oversubscription and correctly allocating resources on HPC systems.
  • OncePerX Types: Introduces OncePerProcess, OncePerThread, and OncePerTask to provide a safe, built-in pattern for code that should only be initialized once within a specific scope.
• BOLT Optimization: Julia can now be built with LLVM's Binary Optimization and Layout Tool (BOLT), which reorders code to improve runtime performance. Benchmarks show improvements of 10-16% on compilation tasks, with up to 23% gains when combined with PGO and LTO.
• Enhanced Atomic Operations: The @atomic macro family now supports array indexing syntax (e.g., @atomic mem[i] += 1), making it easier to perform atomic operations directly on array elements.
• New Package Manager (Pkg) Features:
  • Workspace: Allows multiple projects (e.g., a main package, tests, docs) to share a single manifest file, simplifying dependency management in monorepos.
  • Apps: A new feature to define runnable applications within a package. These can be installed as command-line executables with pre-defined Julia flags, making it easy to distribute CLI tools.
  • Status Highlights: Pkg.status() now visually indicates when a loaded package's version differs from the one specified in the current environment's manifest, helping diagnose version mismatch issues.
• Reproducible Testing with RNG State: When a @testset fails, it now prints the state of its Random Number Generator (RNG). This allows developers to precisely reproduce intermittent test failures for easier debugging.
• Improved LLVM IR Generation: Ptr{T} types are now represented as actual LLVM pointer types instead of integers. This simplifies low-level interoperability with llvmcall and ccall, and removes redundant casting.