Added popcount_u64

4 files changed, 29 insertions, 0 deletions

diff --git a/src/arch/avx2.h b/src/arch/avx2.h
index 5857493..15ce02a 100644
--- a/src/arch/avx2.h
+++ b/src/arch/avx2.h
@@ -34,6 +34,12 @@ popcount_u32(uint32_t x)
         return _mm_popcnt_u32(x);
 }
 
+STATIC_INLINE int
+popcount_u64(uint64_t x)
+{
+        return _mm_popcnt_u64(x);
+}
+
 STATIC void
 pieces(cube_t cube[static 1], uint8_t c[static 8], uint8_t e[static 12])
 {
diff --git a/src/arch/common.h b/src/arch/common.h
index 4412562..2b95fe3 100644
--- a/src/arch/common.h
+++ b/src/arch/common.h
@@ -13,6 +13,7 @@
 #define EFLIP       UINT8_C(0x10)
 
 STATIC_INLINE int popcount_u32(uint32_t);
+STATIC_INLINE int popcount_u64(uint64_t);
 
 STATIC void pieces(cube_t [static 1], uint8_t [static 8], uint8_t [static 12]);
 STATIC_INLINE bool equal(cube_t, cube_t);
diff --git a/src/arch/neon.h b/src/arch/neon.h
index f997813..725e589 100644
--- a/src/arch/neon.h
+++ b/src/arch/neon.h
@@ -47,6 +47,17 @@ popcount_u32(uint32_t x)
         return (int)x;
 }
 
+STATIC_INLINE int
+popcount_u64(uint64_t x)
+{
+        int top, bottom;
+
+        bottom = popcount_u32((uint32_t)(x & UINT64_C(0xFFFFFFFF)));
+        top = popcount_u32((uint32_t)(x >> UINT64_C(32)));
+
+        return top + bottom;
+}
+
 STATIC void
 pieces(cube_t cube[static 1], uint8_t c[static 8], uint8_t e[static 12])
 {
diff --git a/src/arch/portable.h b/src/arch/portable.h
index 4b66fb6..162ed1a 100644
--- a/src/arch/portable.h
+++ b/src/arch/portable.h
@@ -34,6 +34,17 @@ popcount_u32(uint32_t x)
         return (int)x;
 }
 
+STATIC_INLINE int
+popcount_u64(uint64_t x)
+{
+        int top, bottom;
+
+        bottom = popcount_u32((uint32_t)(x & UINT64_C(0xFFFFFFFF)));
+        top = popcount_u32((uint32_t)(x >> UINT64_C(32)));
+
+        return top + bottom;
+}
+
 STATIC void
 pieces(cube_t cube[static 1], uint8_t c[static 8], uint8_t e[static 12])
 {