From d4adcd7ce68aaa3c5828ee665760857b9ff543c4 Mon Sep 17 00:00:00 2001
From: William Hart <goodwillhart@googlemail.com>
Date: Thu, 27 Feb 2014 12:30:53 +0000
Subject: [PATCH] Some very minor speedups/cleanups.

---
 mpn/generic/sb_divappr_q.c | 35 +++++++++++++++++++++++++++++++++--
 mpn/generic/tdiv_q.c       | 12 +++---------
 2 files changed, 36 insertions(+), 11 deletions(-)

diff --git a/mpn/generic/sb_divappr_q.c b/mpn/generic/sb_divappr_q.c
index dc4033bf..d41a4c41 100644
--- a/mpn/generic/sb_divappr_q.c
+++ b/mpn/generic/sb_divappr_q.c
@@ -108,7 +108,7 @@ mpn_sb_divappr_q (mp_ptr qp,
    qn++;
    dp = dp + dn - qn - 1; /* make dp length qn + 1 */
    
-   for ( ; qn > 0; qn--)
+   for ( ; qn > 1; qn--)
      {
        /* fetch next word */
        cy = np[0];
@@ -139,8 +139,39 @@ mpn_sb_divappr_q (mp_ptr qp,
        qp[qn - 1] = q;
        dp++;
      }
+
+     if (qn > 0)
+     {
+       /* fetch next word */
+       cy = np[0];
+ 
+       np--;
+       /* rare case where truncation ruins normalisation */
+       if (cy > dp[1] || (cy == dp[1] && np[0] >= dp[0]))
+         {
+       __divappr_helper(qp, np - 1, dp, 1);
+       return qh;
+         }
+       
+       mpir_divapprox32_preinv2(q, cy, np[0], d1inv);
+         
+       /* np -= dp*q */
+       cy -= mpn_submul_1(np - 1, dp, 2, q);
+
+       /* correct if remainder is too large */
+       if (UNLIKELY(cy || np[0] >= dp[1]))
+         {
+       if (cy || mpn_cmp(np - 1, dp, 2) >= 0)
+         {
+       q++;
+       cy -= mpn_sub_n(np - 1, np - 1, dp, 2);
+         }
+         }
+       
+       qp[0] = q;
+     }
+
      np[1] = cy;
-  
   }
   else
   {
diff --git a/mpn/generic/tdiv_q.c b/mpn/generic/tdiv_q.c
index 22e4ed95..6df70f07 100644
--- a/mpn/generic/tdiv_q.c
+++ b/mpn/generic/tdiv_q.c
@@ -167,8 +167,7 @@ mpn_tdiv_q (mp_ptr qp,
 	}
       else  /* divisor is already normalised */
 	{
-	  if (new_np != np)
-	    MPN_COPY (new_np, np, nn);
+	  MPN_COPY (new_np, np, nn);
 
 	  if (dn == 2)
 	    {
@@ -203,12 +202,7 @@ mpn_tdiv_q (mp_ptr qp,
 
       new_np = scratch;
       new_nn = 2 * qn + 1;
-      if (new_np == np)
-	/* We need {np,nn} to remain untouched until the final adjustment, so
-	   we need to allocate separate space for new_np.  */
-	new_np = TMP_ALLOC_LIMBS (new_nn + 1);
-
-
+      
       dh = dp[dn - 1];
       if (LIKELY ((dh & GMP_NUMB_HIGHBIT) == 0))
 	{
@@ -286,7 +280,7 @@ mpn_tdiv_q (mp_ptr qp,
 	}
 
       MPN_COPY (qp, tp + 1, qn);
-      if (tp[0] <= 4)
+      if (UNLIKELY(tp[0] <= 4))
         {
 	  mp_size_t rn;