Đang tải bảng tin…
Phoronix0 Hot0 bình luận1 phút đọc
One Line x86 Change To GCC Compiler Nets +12% Benchmark Win For Modern Intel/AMD CPUs
Intel software engineer Lili Cui contributed a one-line change to GCC that increases the branch misprediction cost in the generic x86 tuning table. The rationale is that modern CPUs have deeper pipelines, making branch mispredictions more costly, so raising the mispredict scale factor by three helps the compiler avoid pipeline stalls. Benchmarks show a 12.7% gain on Intel Granite Rapids and 12.1% on AMD Zen 5 for the SPEC CPU 2017 544.nab_r NAB test. The change applies to generic x86/x86_64 builds (not -march=native) and has been merged into GCC Git, targeting the GCC 17 stable release.
Nguồn: https://www.phoronix.com/news/GCC-x86-Generic-Mispredict. 8sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.
Đề xuất cho bạn
Hacker News0 Hot15 phút
Lucky Code
A deep dive into how a seemingly cosmetic C code style change — using *lwr++ = x instead of *lwr = x; lwr++ — causes Clang to generate branchless csel instructions on ARM (and cmov on x86), resulting in a 6x speedup for a custom Quicksort implementation. The post shows the full branchless Quicksort with sorting networks, compares the generated assembly with and without branches, and notes that GCC does not exhibit this behavior, consistently generating the slower branch-based version.
