tree 4baaf608a40c275137564eac30e54524f715f172
parent e1d6ee2fab91f83f432493671a434881172c366e
author Matthias Maennich <maennich@google.com> 1623137340 +0100
committer Matthias Maennich <maennich@google.com> 1623149300 +0100

strutil: replace SSE4 specialization with libc call

SkipUntil had been partially specialized to use SSE4.2 instructions if
selected at compile time. The Makefile.ckati sets the compile flags to
the rather incompatible -march=native to enable the SSE4.2 instruction
set. The minimum (Intel) architecture required for SSE4.2 is Nehalem, so
-march=nehalem would do and that would probably make things a bit more
compatible when distributing the binary across machines.

It turns out, the relevant part of SkipUntil can actually be replaced by
a checked `strpbrk` call. Recent libc implementations do specialize
strpbrk for SSE4.2 as well. E.g. glibc runtime-selects `__strpbrk_sse42`
if the host provides the instruction set.

The ckati benchmark `strutil_bench` is neutral to this change. On my
machine the benchmark consistently produces around
   *kati*: WordScanner: 0.156697
with or without the change applied. This is when using the Make based
building process that also sets -march=native.

Other build systems and processes might not want to use -march=native to
be able to produce compatible builds. They might also not want to use
any limiting -march=* at all. That is the case for the Android build
where ckati is built and mainly used. When building in that environment,
the strutil_bench consistently runs
   *kati*: WordScanner: 0.396514
and when applying this patch
   *kati*: WordScanner: 0.255683

The difference between the best Android build and the Make based build
must be related to other build settings done in the Android/Soong based
build, such as FORTIFY_SOURCE, stack protector, etc.

Another side effect of this patch is, that we can run the kati test
suite successfully with --fsanitize=address now. Otherwise the SSE4.2
specific implementation would cause an address violation due to an
out-of-bounds read.

Signed-off-by: Matthias Maennich <maennich@google.com>