DevGex Search

This paper investigates the phenomenon where changing a loop variable from 32-bit unsigned to 64-bit uint64_t causes a 50% performance drop when using the _mm_popcnt_u64 instruction on Intel CPUs. Through assembly analysis and microarchitectural insights, it reveals a false data dependency in the popcnt instruction that propagates across loop iterations, severely limiting instruction-level parallelism. The article details the effects of compiler optimizations, constant vs. non-constant buffer sizes, and the role of the static keyword, providing solutions via inline assembly to break dependency chains. It concludes with best practices for writing high-performance hot loops, emphasizing attention to microarchitectural details and compiler behaviors to avoid such hidden performance pitfalls.

This paper comprehensively examines various algorithms for counting set bits (Hamming Weight) in 32-bit integers. From basic bit-by-bit checking to efficient parallel SWAR algorithms, it provides detailed analysis of Brian Kernighan's algorithm, lookup table methods, and utilization of modern hardware instructions. The article compares performance characteristics of different approaches and offers cross-language implementation examples to help developers choose optimal solutions for specific scenarios.

This article provides an in-depth exploration of various methods for detecting 32-bit and 64-bit architectures in Linux systems, including the use of uname command, analysis of /proc/cpuinfo file, getconf utility, and lshw command. The paper thoroughly examines the principles, applicable scenarios, and limitations of each method, with particular emphasis on the distinction between kernel architecture and CPU architecture. Complete code examples and practical application scenarios are provided, helping developers and system administrators accurately identify system architecture characteristics through systematic comparative analysis.