Streamline criterion for whether to use mpn_lshift or

mpn_rshift.  Increase precision when exp is a multiple of
BITS_PER_MP_LIMB primarily to make exp==0 be a noop.

1 files changed, 22 insertions, 17 deletions

diff --git a/mpf/mul_2exp.c b/mpf/mul_2exp.c
index 5b7345866..6f9e1dc75 100644
--- a/mpf/mul_2exp.c
+++ b/mpf/mul_2exp.c
@@ -51,14 +51,15 @@ mpf_mul_2exp (r, u, exp)
   abs_usize = ABS (usize);
   up = u->_mp_d;
 
-  if (abs_usize > prec)
-    {
-      up += abs_usize - prec;
-      abs_usize = prec;
-    }
-
   if (exp % BITS_PER_MP_LIMB == 0)
     {
+      prec++;			/* retain more precision here as we don't need
+				   to account for carry-out here */
+      if (abs_usize > prec)
+	{
+	  up += abs_usize - prec;
+	  abs_usize = prec;
+	}
       if (rp != up)
 	MPN_COPY_INCR (rp, up, abs_usize);
       r->_mp_exp = uexp + exp / BITS_PER_MP_LIMB;
@@ -66,23 +67,27 @@ mpf_mul_2exp (r, u, exp)
   else
     {
       mp_limb_t cy_limb;
-      if (rp >= up)
+      mp_size_t adj;
+      if (abs_usize > prec)
 	{
-	  cy_limb = mpn_lshift (rp, up, abs_usize, exp % BITS_PER_MP_LIMB);
-	  rp[abs_usize] = cy_limb;
-	  cy_limb = cy_limb != 0;
+	  up += abs_usize - prec;
+	  abs_usize = prec;
+	  /* Use mpn_rshift since mpn_lshift operates downwards, and we
+	     therefore would clobber part of U before using that part, in case
+	     R is the same variable as U.  */
+	  cy_limb = mpn_rshift (rp + 1, up, abs_usize, (-exp) % BITS_PER_MP_LIMB);
+	  rp[0] = cy_limb;
+	  adj = rp[abs_usize] != 0;
 	}
       else
 	{
-	  /* Use mpn_rshift since mpn_lshift operates downwards, and we
-	     therefore would clobber part of U before using that part.  */
-	  cy_limb = mpn_rshift (rp + 1, up, abs_usize, -exp % BITS_PER_MP_LIMB);
-	  rp[0] = cy_limb;
-	  cy_limb = rp[abs_usize] != 0;
+	  cy_limb = mpn_lshift (rp, up, abs_usize, exp % BITS_PER_MP_LIMB);
+	  rp[abs_usize] = cy_limb;
+	  adj = cy_limb != 0;
 	}
 
-      abs_usize += cy_limb;
-      r->_mp_exp = uexp + exp / BITS_PER_MP_LIMB + cy_limb;
+      abs_usize += adj;
+      r->_mp_exp = uexp + exp / BITS_PER_MP_LIMB + adj;
     }
   r->_mp_size = usize >= 0 ? abs_usize : -abs_usize;
 }