Unroll 64-byte CRC32+copy loop for ARMv8

Process 64 bytes per iteration using 8x uint64_t loads
with interleaved memcpy stores and __crc32d calls.
RPi5 benchmarks show 30-51% improvement over the
separate crc32 + memcpy baseline.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

1 files changed, 35 insertions, 0 deletions

diff --git a/arch/arm/crc32_armv8.c b/arch/arm/crc32_armv8.c
index acc6e0be20..59f2b65009 100644
--- a/arch/arm/crc32_armv8.c
+++ b/arch/arm/crc32_armv8.c
@@ -26,6 +26,41 @@ Z_FORCEINLINE static Z_TARGET_CRC uint32_t crc32_copy_impl(uint32_t crc, uint8_t
     if (align_diff)
         c = crc32_armv8_align(c, &dst, &src, &len, align_diff, COPY);
 
+    while (len >= 64) {
+        uint64_t d0 = *(const uint64_t *)src;
+        uint64_t d1 = *(const uint64_t *)(src + 8);
+        uint64_t d2 = *(const uint64_t *)(src + 16);
+        uint64_t d3 = *(const uint64_t *)(src + 24);
+        uint64_t d4 = *(const uint64_t *)(src + 32);
+        uint64_t d5 = *(const uint64_t *)(src + 40);
+        uint64_t d6 = *(const uint64_t *)(src + 48);
+        uint64_t d7 = *(const uint64_t *)(src + 56);
+
+        if (COPY) {
+            memcpy(dst,      &d0, 8);
+            memcpy(dst + 8,  &d1, 8);
+            memcpy(dst + 16, &d2, 8);
+            memcpy(dst + 24, &d3, 8);
+            memcpy(dst + 32, &d4, 8);
+            memcpy(dst + 40, &d5, 8);
+            memcpy(dst + 48, &d6, 8);
+            memcpy(dst + 56, &d7, 8);
+            dst += 64;
+        }
+
+        c = __crc32d(c, d0);
+        c = __crc32d(c, d1);
+        c = __crc32d(c, d2);
+        c = __crc32d(c, d3);
+        c = __crc32d(c, d4);
+        c = __crc32d(c, d5);
+        c = __crc32d(c, d6);
+        c = __crc32d(c, d7);
+
+        src += 64;
+        len -= 64;
+    }
+
     return crc32_armv8_tail(c, dst, src, len, COPY);
 }