We should investigate enabling the read64le feature ( issue 760853 ) for arm/arm64, with measurements to confirm it is a progression.

Using AARCH32 and AARCH64 build of zlib_bench cross-compiled in linux chrome checkout:
  chromium gn args
     target_os = "Android"
     target_cpu = "" || "arm_64"

and the snappy/testdata corpus, on a Pixel C Android Tablet, my initial results are:

  aarch32 build + read64le feature enabled
    - a clear _regression_ decoding zlib wrapped DEFLATE stream
    - gzip wrapped DEFLATE stream, not measured

  aarch64 build + read64le feature enabled
    - _mostly_ a progression decoding zlib wrapped DEFLATE stream
    - gzip wrapped DEFLATE stream, not measured

Looks like we should only consider this feature for ARM64. Work on this bug would be to re-measure on an ARM64 device:

  zlib_bench zlib|gzip ../../snappy/testdata/*
    decoding zlib wrapped DEFLATE stream
    decoding gzip wrapped DEFLATE stream
  image_decode_bench ...
    measure png 140 corpus decoding rate