Hello and welcome to Tech Bytes. Today, we dive into “Load-store conflicts,” an in-depth look at how tiny microarchitectural quirks can wreak havoc on performance. Zachary’s meshoptimizer index decoder saw “significant and unexpected variance in performance” across compilers. Clang-20 hit 6.6 gigabytes per second, but gcc-14’s clever use of SSE raised it to 7.5 GB/s. Then gcc-15 delivered a “devastating performance regression,” plummeting to 4.8 GB/s because separate 32-bit stores confused the store-to-load forwarding unit. On Apple Silicon, clang-17’s paired ldp/stp instructions unlocked almost 10 GB/s by eliminating these store-load conflicts. The takeaway? Even simple code like a 16-element EdgeFifo can stumble over store-load interlocks. For critical loops, inspect your compiler’s generated assembly—and beware of store-load conflicts.
Link to Article