LibCrypto: Use 8-byte crc32 instruction on arm too

Takes

    % time Build/lagom/gunzip -c \
        /Users/thakis/Downloads/trace_bug.json.gz > /dev/null

from 3.9s to 3.87s.

1 files changed, 23 insertions, 4 deletions

diff --git a/Userland/Libraries/LibCrypto/Checksum/CRC32.cpp b/Userland/Libraries/LibCrypto/Checksum/CRC32.cpp
index 805bff2546..7cdddfb274 100644
--- a/Userland/Libraries/LibCrypto/Checksum/CRC32.cpp
+++ b/Userland/Libraries/LibCrypto/Checksum/CRC32.cpp
@@ -13,14 +13,33 @@ namespace Crypto::Checksum {
 
 #if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32)
 
-void CRC32::update(ReadonlyBytes data)
+void CRC32::update(ReadonlyBytes span)
 {
     // FIXME: Does this require runtime checking on rpi?
     //        (Maybe the instruction is present on the rpi4 but not on the rpi3?)
 
-    // FIXME: Use __builtin_arm_crc32d() for aligned middle part.
-    for (size_t i = 0; i < data.size(); i++)
-        m_state = __builtin_arm_crc32b(m_state, data.at(i));
+    u8 const* data = span.data();
+    size_t size = span.size();
+
+    while (size > 0 && (reinterpret_cast<FlatPtr>(data) & 7) != 0) {
+        m_state = __builtin_arm_crc32b(m_state, *data);
+        ++data;
+        --size;
+    }
+
+    auto* data64 = reinterpret_cast<u64 const*>(data);
+    while (size >= 8) {
+        m_state = __builtin_arm_crc32d(m_state, *data64);
+        ++data64;
+        size -= 8;
+    }
+
+    data = reinterpret_cast<u8 const*>(data64);
+    while (size > 0) {
+        m_state = __builtin_arm_crc32b(m_state, *data);
+        ++data;
+        --size;
+    }
 };
 
     // FIXME: On Intel, use _mm_crc32_u8 / _mm_crc32_u64 if available (SSE 4.2).