- Sources: Lattimore post, HN submission
- Summary: David Lattimore, who writes the Wild linker, reproduced the gap mold's benchmarks report and traces most of it to how the benchmark is configured rather than to the linkers themselves. On clang-release the Wild to mold ratio runs from 1.0x on ext4 with the output file deleted and
--no-fork, to 0.6x on tmpfs with the output retained and forking allowed. He states that the tmpfs choice in Wild's own published benchmarks was a mistake because few builds run there, attributes part of the remainder to Wild lacking the fallocate pre-allocation and hugepage mapping mold's author documented and says both are already in Wild for the next release, notes that mold itself got considerably faster in 2.42.0 and 2.42.1 after Wild's August benchmarks were published, and marks the unexplained Threadripper gap as guesswork about thread counts rather than a conclusion. - Why it matters: Linker timings move by close to a factor of two on filesystem and output-file handling alone, so a published linker comparison is unreadable without those settings stated.
send feedback on this story