compress: vectorize ZSTD_count() with SSE2

[rmilkowski]

Summary

Add an SSE2 fast path to ZSTD_count() on x86/x86_64.

After the initial machine-word compare, the new path compares 16 bytes at a time with _mm_cmpeq_epi8() / _mm_movemask_epi8() and uses ZSTD_countTrailingZeros32() to jump to the first mismatch. Non-x86 targets and builds without SSE2 keep the existing scalar path.

Why

ZSTD_count() is a hot helper in match finding. When matches extend past the first word, counting them 16 bytes at a time reduces scalar work on match-heavy inputs.

Benchmark

Host: Intel Xeon Gold 6254, Linux x86_64

Method:

• levels 1..22
• 3 runs per file / level
• page cache warmed before each run

Overall:

• mean compression delta: +16.1%
• median compression delta: +8.3%
• suite CPU time: -1.8%

Largest gains:

• pattern_64kb_gap_0kb_128mb.bin: +98.3% at level 22
• pattern_64kb_gap_0kb_128mb.bin: +86.0% at level 10
• pattern_64kb_gap_0kb_128mb.bin: +75.3% at level 9
• runlength_64kb_pad64_32mb.bin: +54.7% at level 10

Largest regressions observed:

• pattern_1kb_gap_1023kb_128mb.bin: -6.7% at level 8
• high_entropy_190sym_64mb.bin: -4.6% at level 5
• pattern_1kb_gap_1023kb_128mb.bin: -2.7% at level 1

Overall the tradeoff looks favorable: large wins on repetitive / match-heavy inputs, with limited downside on sparse or high-entropy inputs.

Testing

• playTests.sh
• CLI tests
• test-fullbench
• test-fuzzer
• test-zstream
• test-invalidDictionaries
• test-legacy
• test-decodecorpus
• test-pool
• test-longmatch
• CFLAGS='-Werror -O2' make -j1 lib zstd
• make c89build
• make gnu90build
• make c99build
• make cxxtest
• valgrind checks
• CMake configure / build

Notes

• No API change.
• No format change.