Simplify alignment casts in Chorba code

5 files changed, 33 insertions, 21 deletions

diff --git a/arch/generic/crc32_chorba_c.c b/arch/generic/crc32_chorba_c.c
index 87e4fdf674..2ef716de5d 100644
--- a/arch/generic/crc32_chorba_c.c
+++ b/arch/generic/crc32_chorba_c.c
@@ -36,7 +36,7 @@
  * @note Requires minimum input size of 118960 + 512 bytes
  * @note Uses 128KB temporary buffer
  */
-Z_INTERNAL uint32_t crc32_chorba_118960_nondestructive(uint32_t crc, const chorba_word_t *input, size_t len) {
+Z_INTERNAL uint32_t crc32_chorba_118960_nondestructive(uint32_t crc, const uint8_t *buf, size_t len) {
 #if defined(__EMSCRIPTEN__)
     chorba_word_t *bitbuffer = (chorba_word_t*)zng_alloc(bitbuffer_size_bytes);
 #else
@@ -44,7 +44,9 @@ Z_INTERNAL uint32_t crc32_chorba_118960_nondestructive(uint32_t crc, const chorb
 #endif
     const uint8_t *bitbuffer_bytes = (const uint8_t*)bitbuffer;
     uint64a_t *bitbuffer_qwords = (uint64a_t*)bitbuffer;
-    const uint64a_t *input_qwords = (const uint64a_t*)input;
+    /* The calling function ensured that this is aligned correctly */
+    const chorba_word_t* input = (const chorba_word_t*)buf;
+    const uint64a_t* input_qwords = (const uint64a_t*)buf;
 
     size_t i = 0;
 
@@ -486,7 +488,9 @@ Z_INTERNAL uint32_t crc32_chorba_118960_nondestructive(uint32_t crc, const chorb
 
 #  if CHORBA_W == 8
 /* Implement Chorba algorithm from https://arxiv.org/abs/2412.16398 */
-Z_INTERNAL uint32_t crc32_chorba_32768_nondestructive(uint32_t crc, const uint64_t* input, size_t len) {
+Z_INTERNAL uint32_t crc32_chorba_32768_nondestructive(uint32_t crc, const uint8_t* buf, size_t len) {
+    /* The calling function ensured that this is aligned correctly */
+    const uint64_t* input = (const uint64_t*)buf;
     uint64_t bitbuffer[32768 / sizeof(uint64_t)];
     const uint8_t *bitbuffer_bytes = (const uint8_t*)bitbuffer;
     memset(bitbuffer, 0, 32768);
@@ -633,7 +637,9 @@ Z_INTERNAL uint32_t crc32_chorba_32768_nondestructive(uint32_t crc, const uint64
 }
 
 /* Implement Chorba algorithm from https://arxiv.org/abs/2412.16398 */
-Z_INTERNAL uint32_t crc32_chorba_small_nondestructive(uint32_t crc, const uint64_t *input, size_t len) {
+Z_INTERNAL uint32_t crc32_chorba_small_nondestructive(uint32_t crc, const uint8_t *buf, size_t len) {
+    /* The calling function ensured that this is aligned correctly */
+    const uint64_t* input = (const uint64_t*)buf;
     uint64_t final[9] = {0};
     uint64_t next1 = ~crc;
     crc = 0;
@@ -1066,7 +1072,9 @@ Z_INTERNAL uint32_t crc32_chorba_small_nondestructive(uint32_t crc, const uint64
 
 #else // CHORBA_W == 8
 
-Z_INTERNAL uint32_t crc32_chorba_small_nondestructive_32bit(uint32_t crc, const uint32_t *input, size_t len) {
+Z_INTERNAL uint32_t crc32_chorba_small_nondestructive_32bit(uint32_t crc, const uint8_t *buf, size_t len) {
+    /* The calling function ensured that this is aligned correctly */
+    const uint32_t* input = (const uint32_t*)buf;
     uint32_t final[20] = {0};
 
     uint32_t next1 = ~crc;
@@ -1250,13 +1258,13 @@ Z_INTERNAL uint32_t crc32_chorba(uint32_t crc, const uint8_t *buf, size_t len) {
         buf += align_diff;
     }
     if (len > CHORBA_LARGE_THRESHOLD)
-        return crc32_chorba_118960_nondestructive(crc, (const chorba_word_t*)buf, len);
+        return crc32_chorba_118960_nondestructive(crc, buf, len);
 #if CHORBA_W == 8
     if (len > CHORBA_MEDIUM_LOWER_THRESHOLD && len <= CHORBA_MEDIUM_UPPER_THRESHOLD)
-        return crc32_chorba_32768_nondestructive(crc, (const uint64_t*)buf, len);
-    return crc32_chorba_small_nondestructive(crc, (const uint64_t*)buf, len);
+        return crc32_chorba_32768_nondestructive(crc, buf, len);
+    return crc32_chorba_small_nondestructive(crc, buf, len);
 #else
-    return crc32_chorba_small_nondestructive_32bit(crc, (const uint32_t*)buf, len);
+    return crc32_chorba_small_nondestructive_32bit(crc, buf, len);
 #endif
 }
 
diff --git a/arch/x86/crc32_chorba_sse2.c b/arch/x86/crc32_chorba_sse2.c
index 8ecd74443e..66191e046a 100644
--- a/arch/x86/crc32_chorba_sse2.c
+++ b/arch/x86/crc32_chorba_sse2.c
@@ -20,7 +20,9 @@
         d  = _mm_srli_epi64(invec, 20); \
         } while (0);
 
-Z_INTERNAL uint32_t chorba_small_nondestructive_sse2(uint32_t crc, const uint64_t *input, size_t len) {
+Z_INTERNAL uint32_t chorba_small_nondestructive_sse2(uint32_t crc, const uint8_t *buf, size_t len) {
+    /* The calling function ensured that this is aligned correctly */
+    const uint64_t* input = (const uint64_t*)buf;
     ALIGNED_(16) uint64_t final[9] = {0};
     uint64_t next1 = ~crc;
     crc = 0;
@@ -857,9 +859,9 @@ Z_INTERNAL uint32_t crc32_chorba_sse2(uint32_t crc, const uint8_t *buf, size_t l
     }
 #if !defined(WITHOUT_CHORBA)
     if (len > CHORBA_LARGE_THRESHOLD)
-        return crc32_chorba_118960_nondestructive(crc, (const chorba_word_t*)buf, len);
+        return crc32_chorba_118960_nondestructive(crc, buf, len);
 #endif
-    return chorba_small_nondestructive_sse2(crc, (const uint64_t*)buf, len);
+    return chorba_small_nondestructive_sse2(crc, buf, len);
 }
 
 Z_INTERNAL uint32_t crc32_copy_chorba_sse2(uint32_t crc, uint8_t *dst, const uint8_t *src, size_t len) {
diff --git a/arch/x86/crc32_chorba_sse41.c b/arch/x86/crc32_chorba_sse41.c
index 4e750cbd8d..6ef9612440 100644
--- a/arch/x86/crc32_chorba_sse41.c
+++ b/arch/x86/crc32_chorba_sse41.c
@@ -52,7 +52,9 @@
     out3 = _mm_xor_si128(in[4], xor3); \
     } while (0)
 
-Z_FORCEINLINE static uint32_t crc32_chorba_32768_nondestructive_sse41(uint32_t crc, const uint64_t *input, size_t len) {
+Z_FORCEINLINE static uint32_t crc32_chorba_32768_nondestructive_sse41(uint32_t crc, const uint8_t *buf, size_t len) {
+    /* The calling function ensured that this is aligned correctly */
+    const uint64_t* input = (const uint64_t*)buf;
     ALIGNED_(16) uint64_t bitbuffer[32768 / sizeof(uint64_t)];
     __m128i *bitbuffer_v = (__m128i*)bitbuffer;
     const uint8_t *bitbuffer_bytes = (const uint8_t*)bitbuffer;
@@ -315,11 +317,11 @@ Z_INTERNAL uint32_t crc32_chorba_sse41(uint32_t crc, const uint8_t *buf, size_t
     }
 #if !defined(WITHOUT_CHORBA)
     if (len > CHORBA_LARGE_THRESHOLD)
-        return crc32_chorba_118960_nondestructive(crc, (chorba_word_t*)buf, len);
+        return crc32_chorba_118960_nondestructive(crc, buf, len);
 #endif
     if (len > CHORBA_MEDIUM_LOWER_THRESHOLD && len <= CHORBA_MEDIUM_UPPER_THRESHOLD)
-        return crc32_chorba_32768_nondestructive_sse41(crc, (const uint64_t*)buf, len);
-    return chorba_small_nondestructive_sse2(crc, (const uint64_t*)buf, len);
+        return crc32_chorba_32768_nondestructive_sse41(crc, buf, len);
+    return chorba_small_nondestructive_sse2(crc, buf, len);
 }
 
 Z_INTERNAL uint32_t crc32_copy_chorba_sse41(uint32_t crc, uint8_t *dst, const uint8_t *src, size_t len) {
diff --git a/arch/x86/x86_functions.h b/arch/x86/x86_functions.h
index b1d623cace..9162d7eab5 100644
--- a/arch/x86/x86_functions.h
+++ b/arch/x86/x86_functions.h
@@ -25,7 +25,7 @@ void slide_hash_sse2(deflate_state *s);
 #  if !defined(WITHOUT_CHORBA_SSE)
     uint32_t crc32_chorba_sse2(uint32_t crc, const uint8_t *buf, size_t len);
     uint32_t crc32_copy_chorba_sse2(uint32_t crc, uint8_t *dst, const uint8_t *src, size_t len);
-    uint32_t chorba_small_nondestructive_sse2(uint32_t c, const uint64_t *aligned_buf, size_t aligned_len);
+    uint32_t chorba_small_nondestructive_sse2(uint32_t c, const uint8_t *aligned_buf, size_t aligned_len);
 #  endif
 #endif
 
diff --git a/crc32_chorba_p.h b/crc32_chorba_p.h
index f599e707b0..75928a3e58 100644
--- a/crc32_chorba_p.h
+++ b/crc32_chorba_p.h
@@ -26,9 +26,9 @@
     typedef uint32_t chorba_word_t;
 #endif
 
-Z_INTERNAL uint32_t crc32_chorba_118960_nondestructive (uint32_t crc, const chorba_word_t* input, size_t len);
-Z_INTERNAL uint32_t crc32_chorba_32768_nondestructive (uint32_t crc, const uint64_t* input, size_t len);
-Z_INTERNAL uint32_t crc32_chorba_small_nondestructive (uint32_t crc, const uint64_t* input, size_t len);
-Z_INTERNAL uint32_t crc32_chorba_small_nondestructive_32bit (uint32_t crc, const uint32_t* input, size_t len);
+Z_INTERNAL uint32_t crc32_chorba_118960_nondestructive (uint32_t crc, const uint8_t* input, size_t len);
+Z_INTERNAL uint32_t crc32_chorba_32768_nondestructive (uint32_t crc, const uint8_t* input, size_t len);
+Z_INTERNAL uint32_t crc32_chorba_small_nondestructive (uint32_t crc, const uint8_t* input, size_t len);
+Z_INTERNAL uint32_t crc32_chorba_small_nondestructive_32bit (uint32_t crc, const uint8_t* input, size_t len);
 
 #endif /* CRC32_CHORBA_P_H_ */