X-Git-Url: https://git.distorted.org.uk/u/mdw/catacomb/blobdiff_plain/d2d862970aaaa9c14201289f2557bcdfec29b3d5..6d56743beed37f62273209341fa812bcd6aa220f:/mpx-ksqr.c

diff --git a/mpx-ksqr.c b/mpx-ksqr.c
index 226b99a..8a0ad64 100644
--- a/mpx-ksqr.c
+++ b/mpx-ksqr.c
@@ -1,6 +1,6 @@
 /* -*-c-*-
  *
- * $Id: mpx-ksqr.c,v 1.3 2000/06/17 11:42:54 mdw Exp $
+ * $Id: mpx-ksqr.c,v 1.7 2002/10/09 00:36:03 mdw Exp $
  *
  * Karatsuba-based squaring algorithm
  *
@@ -30,6 +30,18 @@
 /*----- Revision history --------------------------------------------------* 
  *
  * $Log: mpx-ksqr.c,v $
+ * Revision 1.7  2002/10/09 00:36:03  mdw
+ * Fix bounds on workspace for Karatsuba operations.
+ *
+ * Revision 1.6  2000/10/08 15:48:35  mdw
+ * Rename Karatsuba constants now that we have @gfx_kmul@ too.
+ *
+ * Revision 1.5  2000/10/08 12:11:01  mdw
+ * Use @mpx_ueq@ instead of @MPX_UCMP@.
+ *
+ * Revision 1.4  2000/07/29 17:04:02  mdw
+ * Remove useless header `mpscan.h'.
+ *
  * Revision 1.3  2000/06/17 11:42:54  mdw
  * Moved the Karatsuba macros into a separate file for better sharing.
  * Fixed some comments.  Use an improved technique so that all the
@@ -49,13 +61,13 @@
 #include <stdio.h>
 
 #include "mpx.h"
-#include "mpx-kmac.h"
+#include "karatsuba.h"
 
 /*----- Tweakables --------------------------------------------------------*/
 
 #ifdef TEST_RIG
-#  undef KARATSUBA_CUTOFF
-#  define KARATSUBA_CUTOFF 2
+#  undef MPK_THRESH
+#  define MPK_THRESH 4
 #endif
 
 /*----- Main code ---------------------------------------------------------*/
@@ -74,9 +86,9 @@
  *		large numbers, although more expensive on small ones, and
  *		rather simpler than full-blown Karatsuba multiplication.
  *
- *		The destination must be twice as large as the argument.  The
- *		scratch space must be twice as large as the argument, plus
- *		the magic number @KARATSUBA_SLOP@.
+ *		The destination must be three times as large as the larger
+ *		argument.  The scratch space must be five times as large as
+ *		the larger argument.
  */
 
 void mpx_ksqr(mpw *dv, mpw *dvl,
@@ -96,7 +108,7 @@ void mpx_ksqr(mpw *dv, mpw *dvl,
 
   MPX_SHRINK(av, avl);
 
-  if (avl - av <= KARATSUBA_CUTOFF) {
+  if (avl - av <= MPK_THRESH) {
     mpx_usqr(dv, dvl, av, avl);
     return;
   }
@@ -117,11 +129,6 @@ 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 --- */
 
   {
@@ -129,13 +136,15 @@ void mpx_ksqr(mpw *dv, mpw *dvl,
     mpw *tdv = dv + m;
     mpw *rdv = tdv + m;
 
+    assert(rdv + m + 4 < dvl);
+    assert(ssv < svl);
     UADD2(sv, svm, av, avm, avm, avl);
-    if (m > KARATSUBA_CUTOFF)
+    if (m > MPK_THRESH)
       mpx_ksqr(tdv, rdv + m + 4, sv, svm + 1, ssv, svl);
     else
       mpx_usqr(tdv, rdv + m + 4, sv, svm + 1);
 
-    if (m > KARATSUBA_CUTOFF)
+    if (m > MPK_THRESH)
       mpx_ksqr(sv, ssv, avm, avl, ssv, svl);
     else
       mpx_usqr(sv, ssv, avm, avl);
@@ -143,7 +152,7 @@ void mpx_ksqr(mpw *dv, mpw *dvl,
     UADD(rdv, sv, svm + 1);
     USUB(tdv, sv, svn);
     
-    if (m > KARATSUBA_CUTOFF)
+    if (m > MPK_THRESH)
       mpx_ksqr(sv, ssv, av, avm, ssv, svl);
     else
       mpx_usqr(sv, ssv, av, avm);
@@ -160,8 +169,6 @@ void mpx_ksqr(mpw *dv, mpw *dvl,
 #include <mLib/alloc.h>
 #include <mLib/testrig.h>
 
-#include "mpscan.h"
-
 #define ALLOC(v, vl, sz) do {                                           \
   size_t _sz = (sz);                                                    \
   mpw *_vv = xmalloc(MPWS(_sz));                                        \
@@ -202,11 +209,11 @@ static int usqr(dstr *v)
   LOAD(a, al, &v[0]);
   LOAD(c, cl, &v[1]);
   m = al - a + 1;
-  ALLOC(d, dl, 2 * m);
-  ALLOC(s, sl, 2 * m + 32);
+  ALLOC(d, dl, 3 * m);
+  ALLOC(s, sl, 5 * m);
 
   mpx_ksqr(d, dl, a, al, s, sl);
-  if (MPX_UCMP(d, dl, !=, c, cl)) {
+  if (!mpx_ueq(d, dl, c, cl)) {
     fprintf(stderr, "\n*** usqr failed\n");
     dumpmp("       a", a, al);
     dumpmp("expected", c, cl);