Simplify and improve.

[u/mdw/catacomb] / mpx-ksqr.c

diff --git a/mpx-ksqr.c b/mpx-ksqr.c
index 45c49c9..3ca145d 100644 (file)
--- 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);
   }
 }