Bilingual

The mask that compiles to nothing: how HotSpots JIT learned to reason about bits
22 days ago
- HotSpot's C2 compiler tracks "known bits" alongside ranges to optimize bitwise operations.
- The optimization eliminates redundant bitwise ANDs, like removing (x << 2) & -4, by proving bits are unchanged.
- The system uses two 32-bit masks per integer: zeros and ones to indicate definite bit states.
- Range and known bits mutually refine each other through a process called canonicalization.
- Rules for bit propagation are defined for operations like AND, OR, and shift, enabling precise analysis.
- The feature was implemented in JDK 26 and improved in JDK 27, with test methods ensuring correctness.
Your code is fast – if you're lucky
16 days ago
- Modern compilers, particularly Clang, optimize loops using branch-free instructions when using certain programming styles.
- A Quicksort implementation with sorting networks and loop unrolling was initially slower than C++ std::sort.
- Rewriting a beginner-friendly if-statement into a compact C idiom (using post-increment operators) significantly improved performance.
- The 'cosmetic' change triggered Clang to replace branches with csel (conditional select) instructions, making the code branchless.
- On x86, Clang similarly generates branchless code using cmov, while GCC consistently produces slower branch-based versions.
- The optimized Quicksort became over 6 times faster than the original and nearly twice as fast as std::sort.
Dense Arena Interning: The Engine of Compiler Performance
15 days ago
- Compiler performance is improved by reducing repeated string and structure checks using Dense Arena Interning.
- Dense Arena Interning converts strings and structures into dense integers during lexing or type construction, enabling O(1) lookups and comparisons in later compiler phases.
- Arena allocators provide stable memory, allowing pointer-based comparisons and efficient memory management without reallocation.
- Dense IDs enable flat array symbol tables, improving cache locality and turning symbol resolution into O(1) array indexing.
- Interning canonicalizes complex types, allowing shallow pointer comparisons instead of deep structural checks, further speeding up type and symbol comparisons.
When Compilers Disagree About UTF‑8
21 hours ago
- An open-source C++ library for UTF-8 handling was initially created without heavy optimization.
- Optimization focused on ASCII shortcut: ASCII characters (range U+0000–U+007F) always pass UTF-8 validity checks, allowing early return.
- A simple ASCII fast-path tripled Clang's performance on pure ASCII text, with 34% improvement on mixed text.
- GCC showed no change on pure ASCII and 3-4% worse performance on mixed text for the same optimization.
- Analysis revealed Clang did not eliminate post-decoding checks, while GCC already optimized them away.
- A refined approach (inline validation in decode functions) improved performance on both compilers: preserved Clang gains, added 15-20% improvement on GCC mixed text.
- Lesson: compiler behavior varies; careful optimization can yield significant gains.

Hasty Briefsbeta

compiler optimization

The mask that compiles to nothing: how HotSpots JIT learned to reason about bits

Your code is fast – if you're lucky

Dense Arena Interning: The Engine of Compiler Performance

When Compilers Disagree About UTF‑8