Fix MOVE_GENERATOR_SSE_ACEPCK build

Author: okuhara

src/bit.h

@@ -3,7 +3,7 @@
  *
  * Bitwise operations header file.
  *
- * @date 1998 - 2025
+ * @date 1998 - 2026
  * @author Richard Delorme
  * @author Toshihiko Okuhara
  * @version 4.5

src/board.h

@@ -182,13 +182,20 @@ extern unsigned char edge_stability[256 * 256];
   #endif
 #endif
 
-#if (MOVE_GENERATOR >= MOVE_GENERATOR_SSE_ACEPCK) && (MOVE_GENERATOR <= MOVE_GENERATOR_AVX512)
+#if (MOVE_GENERATOR >= MOVE_GENERATOR_AVX) && (MOVE_GENERATOR <= MOVE_GENERATOR_AVX512)
 	extern __m128i vectorcall mm_Flip(const __m128i OP, int pos);
 	static inline __m128i vectorcall reduce_vflip(__m128i flip) { return _mm_or_si128(flip, _mm_shuffle_epi32(flip, 0x4e)); }
 	#define	Flip(x,P,O)	((unsigned long long) _mm_cvtsi128_si64(reduce_vflip(mm_Flip(_mm_set_epi64x((O), (P)), (x)))))
 	#define	board_flip(board,x)	((unsigned long long) _mm_cvtsi128_si64(reduce_vflip(mm_Flip(_mm_loadu_si128((__m128i *) (board)), (x)))))
 	#define	vboard_flip(board,x)	((unsigned long long) _mm_cvtsi128_si64(reduce_vflip(mm_Flip((board).v2, (x)))))
 
+#elif MOVE_GENERATOR == MOVE_GENERATOR_SSE_ACEPCK
+	extern __m128i vectorcall mm_Flip(const __m128i OP, int pos);
+	#define reduce_vflip(x)	(x)
+	#define	Flip(x,P,O)	((unsigned long long) _mm_cvtsi128_si64(mm_Flip(_mm_set_epi64x((O), (P)), (x))))
+	#define	board_flip(board,x)	((unsigned long long) _mm_cvtsi128_si64(mm_Flip(_mm_loadu_si128((__m128i *) (board)), (x))))
+	#define	vboard_flip(board,x)	((unsigned long long) _mm_cvtsi128_si64(mm_Flip((board).v2, (x))))
+
 #elif MOVE_GENERATOR == MOVE_GENERATOR_SSE
 	extern __m128i (vectorcall *mm_flip[BOARD_SIZE + 2])(const __m128i);
 	#define	Flip(x,P,O)	((unsigned long long) _mm_cvtsi128_si64(mm_flip[x](_mm_set_epi64x((O), (P)))))
@@ -221,14 +228,13 @@ extern unsigned char edge_stability[256 * 256];
 	extern void init_flip_sse(void);
   #endif
 
-#else
-  #if MOVE_GENERATOR == MOVE_GENERATOR_SSE_BSWAP
+#elif MOVE_GENERATOR == MOVE_GENERATOR_SSE_BSWAP
 	extern unsigned long long Flip(int, unsigned long long, unsigned long long);
-  #else
+	#define	board_flip(board,x)	Flip((x), (board)->player, (board)->opponent)
+
+#else
 	extern unsigned long long (*flip[BOARD_SIZE + 2])(const unsigned long long, const unsigned long long);
 	#define	Flip(x,P,O)	flip[x]((P), (O))
-  #endif
-
 	#define	board_flip(board,x)	Flip((x), (board)->player, (board)->opponent)
 #endif
 

src/flip_sse_acepck.c

@@ -6,7 +6,7 @@
  * Unifiled SSE version using carry propagation by acepck.
  * https://github.com/okuhara/edax-reversi-AVX/issues/5
  *
- * @date 2024
+ * @date 2026
  * @author acepck
  * @author Toshihiko Okuhara
  * @version 4.5
@@ -89,7 +89,7 @@ static const V2DI bb_maska[66][4] = {
  * @param pos player's move.
  * @param P player's disc pattern.
  * @param O opponent's disc pattern.
- * @return flipped disc pattern before reduction.
+ * @return flipped disc pattern.
  */
 
 __m128i vectorcall mm_Flip(__m128i OP, int pos)
@@ -141,5 +141,6 @@ __m128i vectorcall mm_Flip(__m128i OP, int pos)
 		// set flip bits if pos is the break point
 	f3 = _mm_subs_epu8(_mm_and_si128(_mm_add_epi8(OO, f3), _mm_unpacklo_epi64(m34, m34)), f3);
 
-	return _mm_or_si128(_mm_or_si128(_mm_or_si128(f3, f4), _mm_or_si128(f27, f56)), f01);
+	f3 = _mm_or_si128(_mm_or_si128(_mm_or_si128(f3, f4), _mm_or_si128(f27, f56)), f01);
+	return _mm_or_si128(f3, _mm_shuffle_epi32(f3, 0x4e));
 }

src/settings.h

@@ -3,7 +3,7 @@
  *
  * Various macro / constants to control algorithm usage.
  *
- * @date 1998 - 2024
+ * @date 1998 - 2026
  * @author Richard Delorme
  * @version 4.5
  */
@@ -74,6 +74,7 @@
 
 /** SOLID usage */
 #define USE_SOLID true
+
 /** use stability cutoff more often to work best with solid stone */
 #define NWS_STABILITY_SOLID_THRESHOLD 8