Optimizing zlib for ARM

by Adenilson Cavalcanti <​cavalcantii@gmail.com>

• People ship old libpng versions
• libpng, zlib developed around same time
• NEON simd - mandatory in ARMv8
  • checksums easily optimized with simd instructions by checking sums of elements in blocks.
  • adler32, inffast, libpng w/ pallete improved
  • 10% - 10x faster
• work being done to optimize zlib is easier than convincing everyone to upstream their work.
  • especially in projects like chromium, which have many copies of zlib
  • upstreaming optimizations to chromium caused armv6 project to try to use NEON, which broke it
  • security fixes in addition to optimizations exist in chromium's zlib