From c7bd66e7a477115dfecee4a79845a2cadcdfe446 Mon Sep 17 00:00:00 2001
From: "(no author)" <(no author)>
Date: Wed, 3 Mar 2010 00:30:14 +0000
Subject: [PATCH] Finished tuning code for dc_bdiv fns and tuned for K102.

---
 mpn/x86_64/k8/k10/k102/gmp-mparam.h |  2 ++
 tune/common.c                       |  2 +-
 tune/speed.h                        | 20 +++++++-------------
 3 files changed, 10 insertions(+), 14 deletions(-)

diff --git a/mpn/x86_64/k8/k10/k102/gmp-mparam.h b/mpn/x86_64/k8/k10/k102/gmp-mparam.h
index 7d7e0a30..25114c8a 100644
--- a/mpn/x86_64/k8/k10/k102/gmp-mparam.h
+++ b/mpn/x86_64/k8/k10/k102/gmp-mparam.h
@@ -27,6 +27,8 @@
 #define DC_DIVAPPR_Q_N_THRESHOLD        483
 #define INV_DIV_QR_THRESHOLD            400
 #define INV_DIVAPPR_Q_N_THRESHOLD       483
+#define DC_BDIV_QR_THRESHOLD             58
+#define DC_BDIV_Q_THRESHOLD              28
 #define DC_DIVAPPR_Q_THRESHOLD          221
 #define INV_DIVAPPR_Q_THRESHOLD         998
 #define POWM_THRESHOLD                  154
diff --git a/tune/common.c b/tune/common.c
index 5e097629..309682ea 100644
--- a/tune/common.c
+++ b/tune/common.c
@@ -912,7 +912,7 @@ speed_mpn_dc_bdiv_qr_n (struct speed_params *s)
 double
 speed_mpn_dc_bdiv_q (struct speed_params *s)
 {
-  SPEED_ROUTINE_MPN_DC_BDIV_SMALL_Q (mpn_dc_bdiv_q);
+  SPEED_ROUTINE_MPN_DC_BDIV_Q (mpn_dc_bdiv_q);
 }
 
 double
diff --git a/tune/speed.h b/tune/speed.h
index 8ed678f7..e2bb2a47 100644
--- a/tune/speed.h
+++ b/tune/speed.h
@@ -1840,7 +1840,7 @@ int speed_routine_count_zeros_setup _PROTO ((struct speed_params *s,
     return t;								\
   }
 
-#define SPEED_ROUTINE_MPN_DC_BDIV_SMALL_Q(function)				\
+#define SPEED_ROUTINE_MPN_DC_BDIV_Q(function)				\
   {									\
     unsigned  i;							\
     mp_ptr    a, d, q;						\
@@ -1851,22 +1851,18 @@ int speed_routine_count_zeros_setup _PROTO ((struct speed_params *s,
     SPEED_RESTRICT_COND (s->size >= 2);					\
 									\
     TMP_MARK;								\
-    SPEED_TMP_ALLOC_LIMBS (a, 3*s->size, s->align_xp);			\
-    SPEED_TMP_ALLOC_LIMBS (d, 2*s->size,   s->align_yp);			\
+    SPEED_TMP_ALLOC_LIMBS (a, s->size, s->align_xp);			\
+    SPEED_TMP_ALLOC_LIMBS (d, s->size,   s->align_yp);			\
     SPEED_TMP_ALLOC_LIMBS (q, s->size, s->align_wp);			\
     								\
     MPN_COPY (a, s->xp, s->size);					\
-    MPN_COPY (a+s->size, s->xp, s->size);				\
-    MPN_COPY (a+2*s->size, s->xp, s->size);				\
-									\
     MPN_COPY (d, s->yp, s->size);					\
-    MPN_COPY (d+s->size, s->yp, s->size);					\
-									\
+    								\
     /* normalize the data */						\
     d[0] |= 1;					\
     								\
-    speed_operand_src (s, a, 3*s->size);				\
-    speed_operand_src (s, d, 2*s->size);					\
+    speed_operand_src (s, a, s->size);				\
+    speed_operand_src (s, d, s->size);					\
     speed_operand_dst (s, q, s->size);				\
     speed_cache_fill (s);						\
 	                                                         \
@@ -1876,9 +1872,7 @@ int speed_routine_count_zeros_setup _PROTO ((struct speed_params *s,
     i = s->reps;							\
     do	{								\
       MPN_COPY (a, s->xp, s->size);					\
-      MPN_COPY (a+s->size, s->xp, s->size);				\
-      MPN_COPY (a+2*s->size, s->xp, s->size);				\
-      function(q, a, 3*s->size, d, 2*s->size, inv);								\
+      function(q, a, s->size, d, s->size, inv);								\
     } while (--i != 0);							\
     t = speed_endtime ();						\
 									\