Align critical loops to 64 bytes using __align_loops attribute

cmdline/elem.c

@@ -758,6 +758,7 @@ void extent_free(void* void_extent)
 	free(extent);
 }
 
+__aligned(32)
 int extent_parity_compare(const void* void_a, const void* void_b)
 {
 	const struct snapraid_extent* arg_a = void_a;
@@ -771,6 +772,7 @@ int extent_parity_compare(const void* void_a, const void* void_b)
 	return 0;
 }
 
+__aligned(32)
 int extent_file_compare(const void* void_a, const void* void_b)
 {
 	const struct snapraid_extent* arg_a = void_a;

cmdline/elem.h

@@ -1410,6 +1410,7 @@ static inline snapraid_info info_get(tommy_arrayblkof* array, block_off_t pos)
  * Grow the array once and process each internal block directly, avoiding the
  * per-position grow and lookup performed by info_set().
  */
+__aligned_loops(64)
 static inline void info_set_run(tommy_arrayblkof* array, block_off_t pos, block_off_t count, snapraid_info info)
 {
 	tommy_arrayblkof_grow(array, pos + count);

cmdline/state.c

@@ -2294,6 +2294,7 @@ static void decoding_error(const char* path, STREAM* f)
  * pointer directly over cached hashes therefore preserves the same integrity
  * check performed by sread().
  */
+__aligned_loops(64)
 static int state_read_block_run(STREAM* f, struct snapraid_file* file, block_off_t file_pos, block_off_t count, unsigned state, int discard_hash)
 {
 	unsigned char* block_ptr = (unsigned char*)file_block(file, file_pos);
@@ -3902,6 +3903,7 @@ struct state_write_context {
  * copied with a single memcpy(). Keep the stream pointer local while filling
  * each buffered group to avoid the per-block stream call and field updates.
  */
+__aligned_loops(64)
 static int state_write_block_run(STREAM* f, struct snapraid_file* file, block_off_t file_pos, block_off_t count)
 {
 	unsigned char* block_ptr = (unsigned char*)file_block(file, file_pos);

cmdline/util.c

@@ -453,6 +453,7 @@ uint32_t crc32c_gen(uint32_t crc, const unsigned char* ptr, size_t size)
 }
 
 #if CONFIG_X86
+__aligned_loops(64)
 uint32_t crc32c_x86(uint32_t crc, const unsigned char* ptr, size_t size)
 {
 	crc ^= CRC_IV;
@@ -466,6 +467,7 @@ uint32_t crc32c_x86(uint32_t crc, const unsigned char* ptr, size_t size)
 #endif
 
 #if CONFIG_ARM_CRC
+__aligned_loops(64)
 uint32_t crc32c_arm64(uint32_t crc, const unsigned char* ptr, size_t size)
 {
 	crc ^= CRC_IV;

cmdline/util.h

@@ -148,6 +148,7 @@ static __always_inline uint32_t crc32c_x86_64_skip(uint32_t crc, const uint32_t
 	       ^ table[7][crc >> 28];
 }
 
+__aligned_loops(64)
 static __always_inline uint32_t crc32c_x86_64_3lane(uint32_t crc, const unsigned char* ptr, size_t lane_size, const uint32_t table[8][16])
 {
 	const unsigned char* ptr0 = ptr;
@@ -180,6 +181,7 @@ static __always_inline uint32_t crc32c_x86_64_3lane(uint32_t crc, const unsigned
 }
 #endif
 
+__aligned_loops(64)
 static __always_inline uint32_t crc32c_x86_plain(uint32_t crc, const unsigned char* ptr, size_t size)
 {
 #ifdef CONFIG_X86_64

os/portable.h

@@ -70,6 +70,22 @@
 #define __noreturn __attribute__((noreturn))
 #endif
 
+#ifndef __aligned
+#if defined(__GNUC__) || defined(__clang__)
+#define __aligned(n) __attribute__((aligned(n)))
+#else
+#define __aligned(n)
+#endif
+#endif
+
+#ifndef __aligned_loops
+#if defined(__GNUC__) && !defined(__clang__)
+#define __aligned_loops(n) __attribute__((optimize("align-loops=" #n)))
+#else
+#define __aligned_loops(n)
+#endif
+#endif
+
 /**
  * Architecture for inline assembly.
  */