4 files changed, 46 insertions, 0 deletions

diff --git a/src/arch/avx2.h b/src/arch/avx2.h
index 2c3371e..b6ff510 100644
--- a/src/arch/avx2.h
+++ b/src/arch/avx2.h
@@ -16,6 +16,12 @@
 #define SOLVED_CUBE STATIC_CUBE( \
     0, 1, 2, 3, 4, 5, 6, 7, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11)
 
+STATIC_INLINE int
+popcount_u32(uint32_t x)
+{
+        return _mm_popcnt_u32(x);
+}
+
 STATIC void
 pieces(cube_t *cube, uint8_t c[static 8], uint8_t e[static 12])
 {
diff --git a/src/arch/common.h b/src/arch/common.h
index 802790c..52e0fc7 100644
--- a/src/arch/common.h
+++ b/src/arch/common.h
@@ -1,3 +1,5 @@
+STATIC_INLINE int popcount_u32(uint32_t);
+
 STATIC void pieces(cube_t *, uint8_t [static 8], uint8_t [static 12]);
 STATIC_INLINE bool equal(cube_t, cube_t);
 STATIC_INLINE cube_t invertco(cube_t);
@@ -18,6 +20,7 @@ STATIC_INLINE void set_eo(cube_t *, int64_t);
 STATIC_INLINE cube_t invcoord_esep(int64_t);
 
 STATIC_INLINE void invcoord_esep_array(int64_t, int64_t, uint8_t[static 12]);
+STATIC_INLINE cube_t invcoord_eoesep(int64_t);
 
 STATIC_INLINE void
 invcoord_esep_array(int64_t set1, int64_t set2, uint8_t mem[static 12])
@@ -44,3 +47,17 @@ invcoord_esep_array(int64_t set1, int64_t set2, uint8_t mem[static 12])
                 mem[i] = (slice[s]++) | (uint8_t)(s << 2);
         }
 }
+
+STATIC_INLINE cube_t
+invcoord_eoesep(int64_t i)
+{
+        cube_t c;
+        int64_t esep, eo;
+
+        esep = i >> INT64_C(11);
+        eo = i % POW_2_11;
+        c = invcoord_esep(esep);
+        set_eo(&c, eo);
+
+        return c;
+}
diff --git a/src/arch/neon.h b/src/arch/neon.h
index c31c8f9..47dc147 100644
--- a/src/arch/neon.h
+++ b/src/arch/neon.h
@@ -26,6 +26,17 @@ STATIC_INLINE uint8x8_t compose_corners_slim(uint8x8_t, uint8x8_t);
 #define SOLVED_CUBE STATIC_CUBE(        \
         0, 1, 2, 3, 4, 5, 6, 7, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11)
 
+/* TODO: optimize this (use intrinsics?) */
+STATIC_INLINE int
+{
+        int ret;
+
+        for (ret = 0; x != 0; x >>= 1)
+                ret += x & 1;
+
+        return ret;
+}
+
 STATIC void
 pieces(cube_t *cube, uint8_t c[static 8], uint8_t e[static 12])
 {
diff --git a/src/arch/portable.h b/src/arch/portable.h
index e9a8dd0..b8c5211 100644
--- a/src/arch/portable.h
+++ b/src/arch/portable.h
@@ -9,6 +9,18 @@
 #define SOLVED_CUBE STATIC_CUBE( \
     0, 1, 2, 3, 4, 5, 6, 7, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11)
 
+/* TODO: optimize this (use bit tricks?) */
+STATIC_INLINE int
+popcount_u32(uint32_t x)
+{
+        int ret;
+
+        for (ret = 0; x != 0; x >>= 1)
+                ret += x & 1;
+
+        return ret;
+}
+
 STATIC void
 pieces(cube_t *cube, uint8_t c[static 8], uint8_t e[static 12])
 {