commit 2bd100e46361a459b41a29212ea8f97a4837a06c
author Noel Gordon <noel@chromium.org> Sat Apr 30 09:25:00 2022
committer Chromium LUCI CQ <chromium-scoped@luci-project-accounts.iam.gserviceaccount.com> Sat Apr 30 09:25:00 2022
tree e788d7b1e9b7b99d1ae7b6bb3a99c31a9a2f8cbf
parent f293746aa8ff42f89ba514120e2112888fe43fe7

[zlib][asop] Complete USE_ZLIB_RABIN_KARP_ROLLING_HASH feature

If USE_ZLIB_RABIN_KARP_ROLLING_HASH build-time option is defined, use
the Rabin-Karp hash. This disables CRC SIMD hashing on ARM and Intel,
which will degrade compression speed.

However, the compressed output matches canonical zlib output, for the
same input, and that should resolve ASOP OTA issue 1316541.

To ensure the Rabin-Karp hash is used correctly in chromium zlib, the
first step is to go back to using canonical fill_window(). To do this
combine the ARM NEON and Intel SSE2 slide_hash() routines in a common
framework called slide_hash_simd(). Remove fill_window_sse.c and undo
deflate_read_buf() rename: name it back to canonical read_buf().

Change insert_string(): by default it uses CRC32C hashes on all ports
(ARM, Intel) so add code comments to state that. If Rabin-Karp hashes
are enabled, disable CRC32C hashes.

Add a new deflate internal state variable chromium_zlib_hash, used to
detect which type of hashing is enabled (Rabin-Karp, CRC32C). Set the
state variable in deflateInit2_ after cpu_check_features() detection,
with #ifdef guards matching the #ifdef logic of insert_string().

Change canonical fill_window() to insert hashes into {hash,prev} hash
chains based on hash type (Rabin-Karp, CRC32C). Prior to this change,
the ARM port was inserting Rabin-Karp hashes into hash chains even if
CRC32 hashing was active when s->insert was > 0.

Change longest_match() and deflate_fast(): update them to use the new
deflate state variable chromium_zlib_hash.

Compression performance degrades when Rabin-Karp hashing is used, but
is unchanged when CRC32C hashing is enabled (chromium zlib default).

Compat: if Rabin-Karp hashing is enabled, zlib-bench --check built as normal and against canonical zlib 1.2.11, produce the same compressed output for the snappy corpora for gzip,zlib,raw types and compression levels 1..9.

Bug: 1316541
Change-Id: I0d5ee6240f0b7eac4653d60a29d459d994c3871f
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/3596671
Reviewed-by: Chris Blume <cblume@chromium.org>
Commit-Queue: Noel Gordon <noel@chromium.org>
Reviewed-by: Adenilson Cavalcanti <cavalcantii@chromium.org>
Cr-Commit-Position: refs/heads/main@{#998062}