From 4468424ee46294bb2073ca546d54cced2b9bdf02 Mon Sep 17 00:00:00 2001
From: mdw <mdw>
Date: Mon, 13 Dec 1999 15:35:01 +0000
Subject: [PATCH] Simplify and improve.

---
 mpx-kmul.c | 73 +++++++++++++++++++++++++++-----------------------------------
 mpx-ksqr.c | 23 ++++++++++++++------
 2 files changed, 48 insertions(+), 48 deletions(-)

diff --git a/mpx-kmul.c b/mpx-kmul.c
index 7bb4f40..d53622b 100644
--- a/mpx-kmul.c
+++ b/mpx-kmul.c
@@ -1,6 +1,6 @@
 /* -*-c-*-
  *
- * $Id: mpx-kmul.c,v 1.2 1999/12/11 10:58:02 mdw Exp $
+ * $Id: mpx-kmul.c,v 1.3 1999/12/13 15:35:01 mdw Exp $
  *
  * Karatsuba's multiplication algorithm
  *
@@ -30,6 +30,9 @@
 /*----- Revision history --------------------------------------------------* 
  *
  * $Log: mpx-kmul.c,v $
+ * Revision 1.3  1999/12/13 15:35:01  mdw
+ * Simplify and improve.
+ *
  * Revision 1.2  1999/12/11 10:58:02  mdw
  * Remove tweakable comments.
  *
@@ -40,6 +43,7 @@
 
 /*----- Header files ------------------------------------------------------*/
 
+#include <assert.h>
 #include <stdio.h>
 
 #include "mpx.h"
@@ -201,64 +205,51 @@ void mpx_kmul(mpw *dv, mpw *dvl,
       avm = avl;
   }
 
-  /* --- Sort out the middle term --- *
-   *
-   * I'm going to keep track of the carry by hand rather than pass it down to
-   * the next level, because it means multiplication by one or zero, which I
-   * can do easily myself.
-   */
+  assert(((void)"Destination too small for Karatsuba multiply",
+	  dvl - dv >= 4 * m));
+  assert(((void)"Not enough workspace for Karatsuba multiply",
+	  svl - sv >= 4 * m));
+
+  /* --- Sort out the middle term --- */
 
   {
-    unsigned f = 0;
-    enum {
-      carry_a = 1,
-      carry_b = 2
-    };
-
-    mpw *bsv = sv + m, *ssv = bsv + m;
-    mpw *rdv = dv + m, *rdvl = rdv + 2 * m;
-
-    UADD2(sv, bsv + 1, av, avm, avm, avl);
-    if (*bsv)
-      f |= carry_a;
-    UADD2(bsv, ssv + 1, bv, bvm, bvm, bvl);
-    if (*ssv)
-      f |= carry_b;
-    MPX_ZERO(dv, rdv);
+    mpw *bsv = sv + m + 1, *ssv = bsv + m + 1;
+    mpw *rdv = dv + m, *rdvl = rdv + 2 * (m + 2);
+
+    UADD2(sv, bsv, av, avm, avm, avl);
+    UADD2(bsv, ssv, bv, bvm, bvm, bvl);
     if (m > KARATSUBA_CUTOFF)
       mpx_kmul(rdv, rdvl, sv, bsv, bsv, ssv, ssv, svl);
     else
       mpx_umul(rdv, rdvl, sv, bsv, bsv, ssv);
-    MPX_ZERO(rdvl, dvl);
-    rdv += m; rdvl += m;
-    if (f & carry_b)
-      UADD(rdv, sv, bsv);
-    if (f & carry_a)
-      UADD(rdv, bsv, ssv);
-    if (!(~f & (carry_a | carry_b)))
-      MPX_UADDN(rdv + m, rdvl, 1);
   }
 
   /* --- Sort out the other two terms --- */
 
   {
-    mpw *ssv = sv + 2 * m;
+    mpw *svm = sv + m, *svn = svm + m, *ssv = svn + 4;
     mpw *tdv = dv + m;
     mpw *rdv = tdv + m;
 
-    if (m > KARATSUBA_CUTOFF)
-      mpx_kmul(sv, ssv, avm, avl, bvm, bvl, ssv, svl);
-    else
-      mpx_umul(sv, ssv, avm, avl, bvm, bvl);
-    UADD(rdv, sv, ssv);
-    USUB(tdv, sv, ssv);
-    
+    if (avl == avm || bvl == bvm)
+      MPX_ZERO(rdv + m + 1, dvl);
+    else {
+      if (m > KARATSUBA_CUTOFF)
+	mpx_kmul(sv, ssv, avm, avl, bvm, bvl, ssv, svl);
+      else
+	mpx_umul(sv, ssv, avm, avl, bvm, bvl);
+      MPX_COPY(rdv + m + 1, dvl, svm + 1, svn);
+      UADD(rdv, sv, svm + 1);
+      USUB(tdv, sv, svn);
+    }
+
     if (m > KARATSUBA_CUTOFF)
       mpx_kmul(sv, ssv, av, avm, bv, bvm, ssv, svl);
     else
       mpx_umul(sv, ssv, av, avm, bv, bvm);
-    USUB(tdv, sv, ssv);
-    UADD(dv, sv, ssv);
+    MPX_COPY(dv, tdv, sv, svm);
+    USUB(tdv, sv, svn);
+    UADD(tdv, svm, svn);
   }
 }
 
diff --git a/mpx-ksqr.c b/mpx-ksqr.c
index 45c49c9..3ca145d 100644
--- a/mpx-ksqr.c
+++ b/mpx-ksqr.c
@@ -1,6 +1,6 @@
 /* -*-c-*-
  *
- * $Id: mpx-ksqr.c,v 1.1 1999/12/11 10:57:43 mdw Exp $
+ * $Id: mpx-ksqr.c,v 1.2 1999/12/13 15:35:01 mdw Exp $
  *
  * Karatsuba-based squaring algorithm
  *
@@ -30,6 +30,9 @@
 /*----- Revision history --------------------------------------------------* 
  *
  * $Log: mpx-ksqr.c,v $
+ * Revision 1.2  1999/12/13 15:35:01  mdw
+ * Simplify and improve.
+ *
  * Revision 1.1  1999/12/11 10:57:43  mdw
  * Karatsuba squaring algorithm.
  *
@@ -37,6 +40,7 @@
 
 /*----- Header files ------------------------------------------------------*/
 
+#include <assert.h>
 #include <stdio.h>
 
 #include "mpx.h"
@@ -143,10 +147,15 @@ void mpx_ksqr(mpw *dv, mpw *dvl,
   m = (avl - av + 1) >> 1;
   avm = av + m;
 
+  assert(((void)"Destination too small for Karatsuba square",
+	  dvl - dv >= 4 * m));
+  assert(((void)"Not enough workspace for Karatsuba square",
+	  svl - sv >= 4 * m));
+
   /* --- Sort out everything --- */
 
   {
-    mpw *ssv = sv + 2 * m;
+    mpw *svm = sv + m, *svn = svm + m, *ssv = svn + 4;
     mpw *tdv = dv + m;
     mpw *rdv = tdv + m;
 
@@ -160,21 +169,21 @@ void mpx_ksqr(mpw *dv, mpw *dvl,
       mpx_kmul(sv, ssv, av, avm, avm, avl, ssv, svl);
     else
       mpx_umul(sv, ssv, av, avm, avm, avl);
-    ULSL1(tdv, sv, ssv);
-    MPX_ZERO(dv, tdv);
-    MPX_ZERO(rdv + m + 1, dvl);
+    ULSL1(tdv, sv, svn);
 
     if (m > KARATSUBA_CUTOFF)
       mpx_ksqr(sv, ssv, avm, avl, ssv, svl);
     else
       mpx_usqr(sv, ssv, avm, avl);
-    UADD(rdv, sv, ssv);
+    MPX_COPY(rdv + m + 1, dvl, svm + 1, svn);
+    UADD(rdv, sv, svm + 1);
     
     if (m > KARATSUBA_CUTOFF)
       mpx_ksqr(sv, ssv, av, avm, ssv, svl);
     else
       mpx_usqr(sv, ssv, av, avm);
-    UADD(dv, sv, ssv);
+    MPX_COPY(dv, tdv, sv, svm);
+    UADD(tdv, svm, svn);
   }
 }
 
-- 
2.11.0