Improve fd_uint256_cmp

src/ballet/bigint/fd_uint256.h

@@ -5,6 +5,10 @@
 
 #include "../fd_ballet_base.h"
 
+#if FD_HAS_AVX
+#include "../../util/simd/fd_avx.h"
+#endif
+
 /* Align at most at 32 bytes.
    This way a struct containing multiple fd_uint256_t doesn't waste space
    (e.g., on avx512 FD_ALIGNED would be 64, causing each fd_uint256_t to
@@ -52,16 +56,32 @@ fd_uint256_eq( fd_uint256_t const * a,
       && ( a->limbs[3] == b->limbs[3] );
 }
 
-/* fd_uint256_cmp returns 0 is a == b, -1 if a < b, 1 if a > b. */
-static inline int
+/* fd_uint256_cmp returns 0 is a == b, arbitrary negative int if a < b,
+   arbitrary positive int if a > b. */
+
+static inline long
 fd_uint256_cmp( fd_uint256_t const * a,
                 fd_uint256_t const * b ) {
-  for( int i=3; i>=0; i-- ) {
-    if( a->limbs[i] != b->limbs[i] ) {
-      return a->limbs[i] > b->limbs[i] ? 1 : -1;
-    }
-  }
-  return 0;
+# if FD_HAS_AVX && defined(__LZCNT__)
+
+  wu_t va      = wb_ld( a->buf );
+  wu_t vb      = wb_ld( b->buf );
+  uint lt_mask = (uint)_mm256_movemask_epi8( wu_lt( va, vb ) );
+  uint gt_mask = (uint)_mm256_movemask_epi8( wu_gt( va, vb ) );
+  int  lt_prio = (int)( _lzcnt_u32( lt_mask ) );
+  int  gt_prio = (int)( _lzcnt_u32( gt_mask ) );
+  return lt_prio - gt_prio;
+
+# else
+
+  ulong c;
+  ulong d0 = __builtin_subcl( a->limbs[0], b->limbs[0], 0, &c );
+  ulong d1 = __builtin_subcl( a->limbs[1], b->limbs[1], c, &c );
+  ulong d2 = __builtin_subcl( a->limbs[2], b->limbs[2], c, &c );
+  ulong d3 = __builtin_subcl( a->limbs[3], b->limbs[3], c, &c );
+  return c ? -1 : !!(d0|d1|d2|d3);
+
+# endif
 }
 
 /* fd_uint256_bit returns the i-th bit of a.