1 files changed, 61 insertions, 0 deletions
diff --git a/neozip/arch/x86/compare256_avx2.c b/neozip/arch/x86/compare256_avx2.c
new file mode 100644
index 0000000000..5e2b1716cf
--- /dev/null
+++ b/neozip/arch/x86/compare256_avx2.c
@@ -0,0 +1,61 @@
+/* compare256_avx2.c -- AVX2 version of compare256
+ * Copyright Mika T. Lindqvist  <postmaster@raasu.org>
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "zbuild.h"
+#include "zendian.h"
+#include "zmemory.h"
+#include "deflate.h"
+#include "fallback_builtins.h"
+
+#ifdef X86_AVX2
+
+#include <immintrin.h>
+#ifdef _MSC_VER
+#  include <nmmintrin.h>
+#endif
+
+static inline uint32_t compare256_avx2_static(const uint8_t *src0, const uint8_t *src1) {
+    uint32_t len = 0;
+
+    do {
+        __m256i ymm_src0, ymm_src1, ymm_cmp;
+        ymm_src0 = _mm256_loadu_si256((__m256i*)src0);
+        ymm_src1 = _mm256_loadu_si256((__m256i*)src1);
+        ymm_cmp = _mm256_cmpeq_epi8(ymm_src0, ymm_src1); /* non-identical bytes = 00, identical bytes = FF */
+        unsigned mask = (unsigned)_mm256_movemask_epi8(ymm_cmp);
+        if (mask != 0xFFFFFFFF)
+            return len + zng_ctz32(~mask); /* Invert bits so identical = 0 */
+
+        src0 += 32, src1 += 32, len += 32;
+
+        ymm_src0 = _mm256_loadu_si256((__m256i*)src0);
+        ymm_src1 = _mm256_loadu_si256((__m256i*)src1);
+        ymm_cmp = _mm256_cmpeq_epi8(ymm_src0, ymm_src1);
+        mask = (unsigned)_mm256_movemask_epi8(ymm_cmp);
+        if (mask != 0xFFFFFFFF)
+            return len + zng_ctz32(~mask);
+
+        src0 += 32, src1 += 32, len += 32;
+    } while (len < 256);
+
+    return 256;
+}
+
+Z_INTERNAL uint32_t compare256_avx2(const uint8_t *src0, const uint8_t *src1) {
+    return compare256_avx2_static(src0, src1);
+}
+
+#define LONGEST_MATCH       longest_match_avx2
+#define COMPARE256          compare256_avx2_static
+
+#include "match_tpl.h"
+
+#define LONGEST_MATCH_SLOW
+#define LONGEST_MATCH       longest_match_slow_avx2
+#define COMPARE256          compare256_avx2_static
+
+#include "match_tpl.h"
+
+#endif