From 770baa4f555e38bf15ea44308532278612cce457 Mon Sep 17 00:00:00 2001
From: Vladislav Shchapov <vladislav@shchapov.ru>
Date: Thu, 3 Jul 2025 20:58:16 +0500
Subject: [PATCH] Rewrite LoongArch64 CRC32 implementation based on ARMv8 with
 manual alignment

Signed-off-by: Vladislav Shchapov <vladislav@shchapov.ru>
---
 arch/loongarch/crc32_la.c | 68 +++++++++++++++++++++++++++++++--------
 test/test_crc32.cc        |  6 +++-
 2 files changed, 59 insertions(+), 15 deletions(-)

diff --git a/arch/loongarch/crc32_la.c b/arch/loongarch/crc32_la.c
index 4277acbe..034bf7f1 100644
--- a/arch/loongarch/crc32_la.c
+++ b/arch/loongarch/crc32_la.c
@@ -14,26 +14,66 @@
 
 Z_INTERNAL uint32_t crc32_loongarch64(uint32_t crc, const uint8_t *buf,
                                       size_t len) {
-    crc = (~crc) & 0xffffffff;
+    Z_REGISTER uint32_t c;
+    Z_REGISTER uint16_t buf2;
+    Z_REGISTER uint32_t buf4;
+    Z_REGISTER uint64_t buf8;
 
-    while (len >= 8) {
-        crc = (uint32_t)__crc_w_d_w((long int)zng_memread_8(buf), (int)crc);
-        buf += 8;
-        len -= 8;
+    c = ~crc;
+
+    if (UNLIKELY(len == 1)) {
+        c = (uint32_t)__crc_w_b_w((char)(*buf), (int)c);
+        c = ~c;
+        return c;
     }
-    if (len & 4) {
-        crc = (uint32_t)__crc_w_w_w((int)zng_memread_4(buf), (int)crc);
-        buf += 4;
+
+    if ((ptrdiff_t)buf & (sizeof(uint64_t) - 1)) {
+        if (len && ((ptrdiff_t)buf & 1)) {
+            c = (uint32_t)__crc_w_b_w((char)(*buf++), (int)c);
+            len--;
+        }
+
+        if ((len >= sizeof(uint16_t)) && ((ptrdiff_t)buf & (sizeof(uint32_t) - 1))) {
+            buf2 = *((uint16_t*)buf);
+            c = (uint32_t)__crc_w_h_w((short)buf2, (int)c);
+            buf += sizeof(uint16_t);
+            len -= sizeof(uint16_t);
+        }
+
+        if ((len >= sizeof(uint32_t)) && ((ptrdiff_t)buf & (sizeof(uint64_t) - 1))) {
+            buf4 = *((uint32_t*)buf);
+            c = (uint32_t)__crc_w_w_w((int)buf4, (int)c);
+            len -= sizeof(uint32_t);
+            buf += sizeof(uint32_t);
+        }
+
     }
-    if (len & 2) {
-        crc = (uint32_t)__crc_w_h_w((short)zng_memread_2(buf), (int)crc);
-        buf += 2;
+
+    while (len >= sizeof(uint64_t)) {
+        buf8 = *((uint64_t*)buf);
+        c = (uint32_t)__crc_w_d_w((long int)buf8, (int)c);
+        len -= sizeof(uint64_t);
+        buf += sizeof(uint64_t);
     }
-    if (len & 1) {
-        crc = (uint32_t)__crc_w_b_w((char)(*buf), (int)crc);
+
+    if (len & sizeof(uint32_t)) {
+        buf4 = *((uint32_t*)buf);
+        c = (uint32_t)__crc_w_w_w((int)buf4, (int)c);
+        buf += sizeof(uint32_t);
+    }
+
+    if (len & sizeof(uint16_t)) {
+        buf2 = *((uint16_t*)buf);
+        c = (uint32_t)__crc_w_h_w((short)buf2, (int)c);
+        buf += sizeof(uint16_t);
+    }
+
+    if (len & sizeof(uint8_t)) {
+        c = (uint32_t)__crc_w_b_w((char)(*buf), (int)c);
     }
 
-    return crc ^ 0xffffffff;
+    c = ~c;
+    return c;
 }
 
 
diff --git a/test/test_crc32.cc b/test/test_crc32.cc
index b05f220d..d44d079e 100644
--- a/test/test_crc32.cc
+++ b/test/test_crc32.cc
@@ -282,7 +282,7 @@ TEST_CRC32(native, native_crc32, 1)
 
 #else
 
-#ifdef ARM_CRC32
+#if defined(ARM_CRC32) || defined(LOONGARCH_CRC)
 static const int align_offsets[] = {
     1, 2, 3, 4, 5, 6, 7
 };
@@ -295,7 +295,9 @@ static const int align_offsets[] = {
         } \
         hash(GetParam(), func); \
     }
+#endif
 
+#ifdef ARM_CRC32
 INSTANTIATE_TEST_SUITE_P(crc32_alignment, crc32_align, testing::ValuesIn(align_offsets));
 TEST_CRC32(armv8, crc32_armv8, test_cpu_features.arm.has_crc32)
 TEST_CRC32_ALIGN(armv8_align, crc32_armv8, test_cpu_features.arm.has_crc32)
@@ -322,7 +324,9 @@ TEST_CRC32(chorba_sse2, crc32_chorba_sse2, test_cpu_features.x86.has_sse2)
 TEST_CRC32(chorba_sse41, crc32_chorba_sse41, test_cpu_features.x86.has_sse41)
 #endif
 #if defined(LOONGARCH_CRC)
+INSTANTIATE_TEST_SUITE_P(crc32_alignment, crc32_align, testing::ValuesIn(align_offsets));
 TEST_CRC32(loongarch64, crc32_loongarch64, test_cpu_features.loongarch.has_crc)
+TEST_CRC32_ALIGN(loongarch64_align, crc32_loongarch64, test_cpu_features.loongarch.has_crc)
 #endif
 
 #endif
-- 
2.47.2