X-Git-Url: https://git.openssl.org/?p=openssl.git;a=blobdiff_plain;f=crypto%2Fbn%2Fbn_mul.c;h=eb5d5256137ea5861f3cf71d2f07ae83b75afbe9;hp=3e8d8b9567a89af94383b8e6120a78cb060a55a5;hb=90e7ce363ae8278bb4ac647739c89a43ed341e34;hpb=47ff5de8c2efa09392dab257c4616881ba3b8037

diff --git a/crypto/bn/bn_mul.c b/crypto/bn/bn_mul.c
index 3e8d8b9567..eb5d525613 100644
--- a/crypto/bn/bn_mul.c
+++ b/crypto/bn/bn_mul.c
@@ -56,10 +56,323 @@
  * [including the GNU Public Licence.]
  */
 
+#ifndef BN_DEBUG
+# undef NDEBUG /* avoid conflicting definitions */
+# define NDEBUG
+#endif
+
 #include <stdio.h>
+#include <assert.h>
 #include "cryptlib.h"
 #include "bn_lcl.h"
 
+/* Here follows specialised variants of bn_add_words() and
+   bn_sub_words().  They have the property performing operations on
+   arrays of different sizes.  The sizes of those arrays is expressed through
+   cl, which is the common length ( basicall, min(len(a),len(b)) ), and dl,
+   which is the delta between the two lengths, calculated as len(a)-len(b).
+   All lengths are the number of BN_ULONGs...  For the operations that require
+   a result array as parameter, it must have the length cl+abs(dl).
+   These functions should probably end up in bn_asm.c as soon as there are
+   assembler counterparts for the systems that use assembler files.  */
+
+BN_ULONG bn_sub_part_words(BN_ULONG *r,
+	const BN_ULONG *a, const BN_ULONG *b,
+	int cl, int dl)
+	{
+	BN_ULONG c, t;
+
+	assert(cl >= 0);
+	c = bn_sub_words(r, a, b, cl);
+
+	if (dl == 0)
+		return c;
+
+	r += cl;
+	a += cl;
+	b += cl;
+
+	if (dl < 0)
+		{
+#ifdef BN_COUNT
+		fprintf(stderr, "  bn_sub_part_words %d + %d (dl < 0, c = %d)\n", cl, dl, c);
+#endif
+		for (;;)
+			{
+			t = b[0];
+			r[0] = (0-t-c)&BN_MASK2;
+			if (t != 0) c=1;
+			if (++dl >= 0) break;
+
+			t = b[1];
+			r[1] = (0-t-c)&BN_MASK2;
+			if (t != 0) c=1;
+			if (++dl >= 0) break;
+
+			t = b[2];
+			r[2] = (0-t-c)&BN_MASK2;
+			if (t != 0) c=1;
+			if (++dl >= 0) break;
+
+			t = b[3];
+			r[3] = (0-t-c)&BN_MASK2;
+			if (t != 0) c=1;
+			if (++dl >= 0) break;
+
+			b += 4;
+			r += 4;
+			}
+		}
+	else
+		{
+		int save_dl = dl;
+#ifdef BN_COUNT
+		fprintf(stderr, "  bn_sub_part_words %d + %d (dl > 0, c = %d)\n", cl, dl, c);
+#endif
+		while(c)
+			{
+			t = a[0];
+			r[0] = (t-c)&BN_MASK2;
+			if (t != 0) c=0;
+			if (--dl <= 0) break;
+
+			t = a[1];
+			r[1] = (t-c)&BN_MASK2;
+			if (t != 0) c=0;
+			if (--dl <= 0) break;
+
+			t = a[2];
+			r[2] = (t-c)&BN_MASK2;
+			if (t != 0) c=0;
+			if (--dl <= 0) break;
+
+			t = a[3];
+			r[3] = (t-c)&BN_MASK2;
+			if (t != 0) c=0;
+			if (--dl <= 0) break;
+
+			save_dl = dl;
+			a += 4;
+			r += 4;
+			}
+		if (dl > 0)
+			{
+#ifdef BN_COUNT
+			fprintf(stderr, "  bn_sub_part_words %d + %d (dl > 0, c == 0)\n", cl, dl);
+#endif
+			if (save_dl > dl)
+				{
+				switch (save_dl - dl)
+					{
+				case 1:
+					r[1] = a[1];
+					if (--dl <= 0) break;
+				case 2:
+					r[2] = a[2];
+					if (--dl <= 0) break;
+				case 3:
+					r[3] = a[3];
+					if (--dl <= 0) break;
+					}
+				a += 4;
+				r += 4;
+				}
+			}
+		if (dl > 0)
+			{
+#ifdef BN_COUNT
+			fprintf(stderr, "  bn_sub_part_words %d + %d (dl > 0, copy)\n", cl, dl);
+#endif
+			for(;;)
+				{
+				r[0] = a[0];
+				if (--dl <= 0) break;
+				r[1] = a[1];
+				if (--dl <= 0) break;
+				r[2] = a[2];
+				if (--dl <= 0) break;
+				r[3] = a[3];
+				if (--dl <= 0) break;
+
+				a += 4;
+				r += 4;
+				}
+			}
+		}
+	return c;
+	}
+
+BN_ULONG bn_add_part_words(BN_ULONG *r,
+	const BN_ULONG *a, const BN_ULONG *b,
+	int cl, int dl)
+	{
+	BN_ULONG c, l, t;
+
+	assert(cl >= 0);
+	c = bn_add_words(r, a, b, cl);
+
+	if (dl == 0)
+		return c;
+
+	r += cl;
+	a += cl;
+	b += cl;
+
+	if (dl < 0)
+		{
+		int save_dl = dl;
+#ifdef BN_COUNT
+		fprintf(stderr, "  bn_add_part_words %d + %d (dl < 0, c = %d)\n", cl, dl, c);
+#endif
+		while (c)
+			{
+			l=(c+b[0])&BN_MASK2;
+			c=(l < c);
+			r[0]=l;
+			if (++dl >= 0) break;
+
+			l=(c+b[1])&BN_MASK2;
+			c=(l < c);
+			r[1]=l;
+			if (++dl >= 0) break;
+
+			l=(c+b[2])&BN_MASK2;
+			c=(l < c);
+			r[2]=l;
+			if (++dl >= 0) break;
+
+			l=(c+b[3])&BN_MASK2;
+			c=(l < c);
+			r[3]=l;
+			if (++dl >= 0) break;
+
+			save_dl = dl;
+			b+=4;
+			r+=4;
+			}
+		if (dl < 0)
+			{
+#ifdef BN_COUNT
+			fprintf(stderr, "  bn_add_part_words %d + %d (dl < 0, c == 0)\n", cl, dl);
+#endif
+			if (save_dl < dl)
+				{
+				switch (dl - save_dl)
+					{
+				case 1:
+					r[1] = b[1];
+					if (++dl >= 0) break;
+				case 2:
+					r[2] = b[2];
+					if (++dl >= 0) break;
+				case 3:
+					r[3] = b[3];
+					if (++dl >= 0) break;
+					}
+				b += 4;
+				r += 4;
+				}
+			}
+		if (dl < 0)
+			{
+#ifdef BN_COUNT
+			fprintf(stderr, "  bn_add_part_words %d + %d (dl < 0, copy)\n", cl, dl);
+#endif
+			for(;;)
+				{
+				r[0] = b[0];
+				if (++dl >= 0) break;
+				r[1] = b[1];
+				if (++dl >= 0) break;
+				r[2] = b[2];
+				if (++dl >= 0) break;
+				r[3] = b[3];
+				if (++dl >= 0) break;
+
+				b += 4;
+				r += 4;
+				}
+			}
+		}
+	else
+		{
+		int save_dl = dl;
+#ifdef BN_COUNT
+		fprintf(stderr, "  bn_add_part_words %d + %d (dl > 0)\n", cl, dl);
+#endif
+		while (c)
+			{
+			t=(a[0]+c)&BN_MASK2;
+			c=(t < c);
+			r[0]=t;
+			if (--dl <= 0) break;
+
+			t=(a[1]+c)&BN_MASK2;
+			c=(t < c);
+			r[1]=t;
+			if (--dl <= 0) break;
+
+			t=(a[2]+c)&BN_MASK2;
+			c=(t < c);
+			r[2]=t;
+			if (--dl <= 0) break;
+
+			t=(a[3]+c)&BN_MASK2;
+			c=(t < c);
+			r[3]=t;
+			if (--dl <= 0) break;
+
+			save_dl = dl;
+			a+=4;
+			r+=4;
+			}
+#ifdef BN_COUNT
+		fprintf(stderr, "  bn_add_part_words %d + %d (dl > 0, c == 0)\n", cl, dl);
+#endif
+		if (dl > 0)
+			{
+			if (save_dl > dl)
+				{
+				switch (save_dl - dl)
+					{
+				case 1:
+					r[1] = a[1];
+					if (--dl <= 0) break;
+				case 2:
+					r[2] = a[2];
+					if (--dl <= 0) break;
+				case 3:
+					r[3] = a[3];
+					if (--dl <= 0) break;
+					}
+				a += 4;
+				r += 4;
+				}
+			}
+		if (dl > 0)
+			{
+#ifdef BN_COUNT
+			fprintf(stderr, "  bn_add_part_words %d + %d (dl > 0, copy)\n", cl, dl);
+#endif
+			for(;;)
+				{
+				r[0] = a[0];
+				if (--dl <= 0) break;
+				r[1] = a[1];
+				if (--dl <= 0) break;
+				r[2] = a[2];
+				if (--dl <= 0) break;
+				r[3] = a[3];
+				if (--dl <= 0) break;
+
+				a += 4;
+				r += 4;
+				}
+			}
+		}
+	return c;
+	}
+
 #ifdef BN_RECURSION
 /* Karatsuba recursive multiplication algorithm
  * (cf. Knuth, The Art of Computer Programming, Vol. 2) */
@@ -75,14 +388,15 @@
  * a[1]*b[1]
  */
 void bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
-	     BN_ULONG *t)
+	int dna, int dnb, BN_ULONG *t)
 	{
 	int n=n2/2,c1,c2;
+	int tna=n+dna, tnb=n+dnb;
 	unsigned int neg,zero;
 	BN_ULONG ln,lo,*p;
 
 # ifdef BN_COUNT
-	printf(" bn_mul_recursive %d * %d\n",n2,n2);
+	fprintf(stderr," bn_mul_recursive %d * %d\n",n2,n2);
 # endif
 # ifdef BN_MUL_COMBA
 #  if 0
@@ -105,21 +419,21 @@ void bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
 		return;
 		}
 	/* r=(a[0]-a[1])*(b[1]-b[0]) */
-	c1=bn_cmp_words(a,&(a[n]),n);
-	c2=bn_cmp_words(&(b[n]),b,n);
+	c1=bn_cmp_part_words(a,&(a[n]),tna,n-tna);
+	c2=bn_cmp_part_words(&(b[n]),b,tnb,tnb-n);
 	zero=neg=0;
 	switch (c1*3+c2)
 		{
 	case -4:
-		bn_sub_words(t,      &(a[n]),a,      n); /* - */
-		bn_sub_words(&(t[n]),b,      &(b[n]),n); /* - */
+		bn_sub_part_words(t,      &(a[n]),a,      tna,tna-n); /* - */
+		bn_sub_part_words(&(t[n]),b,      &(b[n]),tnb,n-tnb); /* - */
 		break;
 	case -3:
 		zero=1;
 		break;
 	case -2:
-		bn_sub_words(t,      &(a[n]),a,      n); /* - */
-		bn_sub_words(&(t[n]),&(b[n]),b,      n); /* + */
+		bn_sub_part_words(t,      &(a[n]),a,      tna,tna-n); /* - */
+		bn_sub_part_words(&(t[n]),&(b[n]),b,      tnb,tnb-n); /* + */
 		neg=1;
 		break;
 	case -1:
@@ -128,16 +442,16 @@ void bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
 		zero=1;
 		break;
 	case 2:
-		bn_sub_words(t,      a,      &(a[n]),n); /* + */
-		bn_sub_words(&(t[n]),b,      &(b[n]),n); /* - */
+		bn_sub_part_words(t,      a,      &(a[n]),tna,n-tna); /* + */
+		bn_sub_part_words(&(t[n]),b,      &(b[n]),tnb,n-tnb); /* - */
 		neg=1;
 		break;
 	case 3:
 		zero=1;
 		break;
 	case 4:
-		bn_sub_words(t,      a,      &(a[n]),n);
-		bn_sub_words(&(t[n]),&(b[n]),b,      n);
+		bn_sub_part_words(t,      a,      &(a[n]),tna,n-tna);
+		bn_sub_part_words(&(t[n]),&(b[n]),b,      tnb,tnb-n);
 		break;
 		}
 
@@ -167,11 +481,11 @@ void bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
 		{
 		p= &(t[n2*2]);
 		if (!zero)
-			bn_mul_recursive(&(t[n2]),t,&(t[n]),n,p);
+			bn_mul_recursive(&(t[n2]),t,&(t[n]),n,0,0,p);
 		else
 			memset(&(t[n2]),0,n2*sizeof(BN_ULONG));
-		bn_mul_recursive(r,a,b,n,p);
-		bn_mul_recursive(&(r[n2]),&(a[n]),&(b[n]),n,p);
+		bn_mul_recursive(r,a,b,n,0,0,p);
+		bn_mul_recursive(&(r[n2]),&(a[n]),&(b[n]),n,dna,dnb,p);
 		}
 
 	/* t[32] holds (a[0]-a[1])*(b[1]-b[0]), c1 is the sign
@@ -220,39 +534,39 @@ void bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
 
 /* n+tn is the word length
  * t needs to be n*4 is size, as does r */
-void bn_mul_part_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int tn,
-	     int n, BN_ULONG *t)
+void bn_mul_part_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n,
+	     int tna, int tnb, BN_ULONG *t)
 	{
 	int i,j,n2=n*2;
 	unsigned int c1,c2,neg,zero;
 	BN_ULONG ln,lo,*p;
 
 # ifdef BN_COUNT
-	printf(" bn_mul_part_recursive %d * %d\n",tn+n,tn+n);
+	fprintf(stderr," bn_mul_part_recursive (%d+%d) * (%d+%d)\n",
+		tna, n, tnb, n);
 # endif
 	if (n < 8)
 		{
-		i=tn+n;
-		bn_mul_normal(r,a,i,b,i);
+		bn_mul_normal(r,a,n+tna,b,n+tnb);
 		return;
 		}
 
 	/* r=(a[0]-a[1])*(b[1]-b[0]) */
-	c1=bn_cmp_words(a,&(a[n]),n);
-	c2=bn_cmp_words(&(b[n]),b,n);
+	c1=bn_cmp_part_words(a,&(a[n]),tna,n-tna);
+	c2=bn_cmp_part_words(&(b[n]),b,tnb,tnb-n);
 	zero=neg=0;
 	switch (c1*3+c2)
 		{
 	case -4:
-		bn_sub_words(t,      &(a[n]),a,      n); /* - */
-		bn_sub_words(&(t[n]),b,      &(b[n]),n); /* - */
+		bn_sub_part_words(t,      &(a[n]),a,      tna,tna-n); /* - */
+		bn_sub_part_words(&(t[n]),b,      &(b[n]),tnb,n-tnb); /* - */
 		break;
 	case -3:
 		zero=1;
 		/* break; */
 	case -2:
-		bn_sub_words(t,      &(a[n]),a,      n); /* - */
-		bn_sub_words(&(t[n]),&(b[n]),b,      n); /* + */
+		bn_sub_part_words(t,      &(a[n]),a,      tna,tna-n); /* - */
+		bn_sub_part_words(&(t[n]),&(b[n]),b,      tnb,tnb-n); /* + */
 		neg=1;
 		break;
 	case -1:
@@ -261,16 +575,16 @@ void bn_mul_part_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int tn,
 		zero=1;
 		/* break; */
 	case 2:
-		bn_sub_words(t,      a,      &(a[n]),n); /* + */
-		bn_sub_words(&(t[n]),b,      &(b[n]),n); /* - */
+		bn_sub_part_words(t,      a,      &(a[n]),tna,n-tna); /* + */
+		bn_sub_part_words(&(t[n]),b,      &(b[n]),tnb,n-tnb); /* - */
 		neg=1;
 		break;
 	case 3:
 		zero=1;
 		/* break; */
 	case 4:
-		bn_sub_words(t,      a,      &(a[n]),n);
-		bn_sub_words(&(t[n]),&(b[n]),b,      n);
+		bn_sub_part_words(t,      a,      &(a[n]),tna,n-tna);
+		bn_sub_part_words(&(t[n]),&(b[n]),b,      tnb,tnb-n);
 		break;
 		}
 		/* The zero case isn't yet implemented here. The speedup
@@ -289,54 +603,59 @@ void bn_mul_part_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int tn,
 		{
 		bn_mul_comba8(&(t[n2]),t,&(t[n]));
 		bn_mul_comba8(r,a,b);
-		bn_mul_normal(&(r[n2]),&(a[n]),tn,&(b[n]),tn);
-		memset(&(r[n2+tn*2]),0,sizeof(BN_ULONG)*(n2-tn*2));
+		bn_mul_normal(&(r[n2]),&(a[n]),tna,&(b[n]),tnb);
+		memset(&(r[n2+tna+tnb]),0,sizeof(BN_ULONG)*(n2-tna-tnb));
 		}
 	else
 		{
 		p= &(t[n2*2]);
-		bn_mul_recursive(&(t[n2]),t,&(t[n]),n,p);
-		bn_mul_recursive(r,a,b,n,p);
+		bn_mul_recursive(&(t[n2]),t,&(t[n]),n,0,0,p);
+		bn_mul_recursive(r,a,b,n,0,0,p);
 		i=n/2;
 		/* If there is only a bottom half to the number,
 		 * just do it */
-		j=tn-i;
+		if (tna > tnb)
+			j = tna - i;
+		else
+			j = tnb - i;
 		if (j == 0)
 			{
-			bn_mul_recursive(&(r[n2]),&(a[n]),&(b[n]),i,p);
+			bn_mul_recursive(&(r[n2]),&(a[n]),&(b[n]),
+				i,tna-i,tnb-i,p);
 			memset(&(r[n2+i*2]),0,sizeof(BN_ULONG)*(n2-i*2));
 			}
 		else if (j > 0) /* eg, n == 16, i == 8 and tn == 11 */
 				{
 				bn_mul_part_recursive(&(r[n2]),&(a[n]),&(b[n]),
-					j,i,p);
-				memset(&(r[n2+tn*2]),0,
-					sizeof(BN_ULONG)*(n2-tn*2));
+					i,tna-i,tnb-i,p);
+				memset(&(r[n2+tna+tnb]),0,
+					sizeof(BN_ULONG)*(n2-tna-tnb));
 				}
 		else /* (j < 0) eg, n == 16, i == 8 and tn == 5 */
 			{
 			memset(&(r[n2]),0,sizeof(BN_ULONG)*n2);
-			if (tn < BN_MUL_RECURSIVE_SIZE_NORMAL)
+			if (tna < BN_MUL_RECURSIVE_SIZE_NORMAL
+				&& tnb < BN_MUL_RECURSIVE_SIZE_NORMAL)
 				{
-				bn_mul_normal(&(r[n2]),&(a[n]),tn,&(b[n]),tn);
+				bn_mul_normal(&(r[n2]),&(a[n]),tna,&(b[n]),tnb);
 				}
 			else
 				{
 				for (;;)
 					{
 					i/=2;
-					if (i < tn)
+					if (i < tna && i < tnb)
 						{
 						bn_mul_part_recursive(&(r[n2]),
 							&(a[n]),&(b[n]),
-							tn-i,i,p);
+							i,tna-i,tnb-i,p);
 						break;
 						}
-					else if (i == tn)
+					else if (i <= tna && i <= tnb)
 						{
 						bn_mul_recursive(&(r[n2]),
 							&(a[n]),&(b[n]),
-							i,p);
+							i,tna-i,tnb-i,p);
 						break;
 						}
 					}
@@ -397,10 +716,10 @@ void bn_mul_low_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
 	int n=n2/2;
 
 # ifdef BN_COUNT
-	printf(" bn_mul_low_recursive %d * %d\n",n2,n2);
+	fprintf(stderr," bn_mul_low_recursive %d * %d\n",n2,n2);
 # endif
 
-	bn_mul_recursive(r,a,b,n,&(t[0]));
+	bn_mul_recursive(r,a,b,n,0,0,&(t[0]));
 	if (n >= BN_MUL_LOW_RECURSIVE_SIZE_NORMAL)
 		{
 		bn_mul_low_recursive(&(t[0]),&(a[0]),&(b[n]),n,&(t[n2]));
@@ -431,7 +750,7 @@ void bn_mul_high(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, BN_ULONG *l, int n2,
 	BN_ULONG ll,lc,*lp,*mp;
 
 # ifdef BN_COUNT
-	printf(" bn_mul_high %d * %d\n",n2,n2);
+	fprintf(stderr," bn_mul_high %d * %d\n",n2,n2);
 # endif
 	n=n2/2;
 
@@ -484,8 +803,8 @@ void bn_mul_high(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, BN_ULONG *l, int n2,
 	else
 # endif
 		{
-		bn_mul_recursive(&(t[0]),&(r[0]),&(r[n]),n,&(t[n2]));
-		bn_mul_recursive(r,&(a[n]),&(b[n]),n,&(t[n2]));
+		bn_mul_recursive(&(t[0]),&(r[0]),&(r[n]),n,0,0,&(t[n2]));
+		bn_mul_recursive(r,&(a[n]),&(b[n]),n,0,0,&(t[n2]));
 		}
 
 	/* s0 == low(al*bl)
@@ -608,11 +927,11 @@ void bn_mul_high(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, BN_ULONG *l, int n2,
 	}
 #endif /* BN_RECURSION */
 
-int BN_mul(BIGNUM *r, BIGNUM *a, BIGNUM *b, BN_CTX *ctx)
+int BN_mul(BIGNUM *r, const BIGNUM *a, const BIGNUM *b, BN_CTX *ctx)
 	{
+	int ret=0;
 	int top,al,bl;
 	BIGNUM *rr;
-	int ret = 0;
 #if defined(BN_MUL_COMBA) || defined(BN_RECURSION)
 	int i;
 #endif
@@ -622,7 +941,7 @@ int BN_mul(BIGNUM *r, BIGNUM *a, BIGNUM *b, BN_CTX *ctx)
 #endif
 
 #ifdef BN_COUNT
-	printf("BN_mul %d * %d\n",a->top,b->top);
+	fprintf(stderr,"BN_mul %d * %d\n",a->top,b->top);
 #endif
 
 	bn_check_top(a);
@@ -675,17 +994,55 @@ int BN_mul(BIGNUM *r, BIGNUM *a, BIGNUM *b, BN_CTX *ctx)
 #ifdef BN_RECURSION
 	if ((al >= BN_MULL_SIZE_NORMAL) && (bl >= BN_MULL_SIZE_NORMAL))
 		{
+		if (i >= -1 && i <= 1)
+			{
+			int sav_j =0;
+			/* Find out the power of two lower or equal
+			   to the longest of the two numbers */
+			if (i >= 0)
+				{
+				j = BN_num_bits_word((BN_ULONG)al);
+				}
+			if (i == -1)
+				{
+				j = BN_num_bits_word((BN_ULONG)bl);
+				}
+			sav_j = j;
+			j = 1<<(j-1);
+			assert(j <= al || j <= bl);
+			k = j+j;
+			t = BN_CTX_get(ctx);
+			if (al > j || bl > j)
+				{
+				bn_wexpand(t,k*4);
+				bn_wexpand(rr,k*4);
+				bn_mul_part_recursive(rr->d,a->d,b->d,
+					j,al-j,bl-j,t->d);
+				}
+			else	/* al <= j || bl <= j */
+				{
+				bn_wexpand(t,k*2);
+				bn_wexpand(rr,k*2);
+				bn_mul_recursive(rr->d,a->d,b->d,
+					j,al-j,bl-j,t->d);
+				}
+			rr->top=top;
+			goto end;
+			}
+#if 0
 		if (i == 1 && !BN_get_flags(b,BN_FLG_STATIC_DATA))
 			{
-			bn_wexpand(b,al);
-			b->d[bl]=0;
+			BIGNUM *tmp_bn = (BIGNUM *)b;
+			bn_wexpand(tmp_bn,al);
+			tmp_bn->d[bl]=0;
 			bl++;
 			i--;
 			}
 		else if (i == -1 && !BN_get_flags(a,BN_FLG_STATIC_DATA))
 			{
-			bn_wexpand(a,bl);
-			a->d[al]=0;
+			BIGNUM *tmp_bn = (BIGNUM *)a;
+			bn_wexpand(tmp_bn,bl);
+			tmp_bn->d[al]=0;
 			al++;
 			i++;
 			}
@@ -705,19 +1062,14 @@ int BN_mul(BIGNUM *r, BIGNUM *a, BIGNUM *b, BN_CTX *ctx)
 				}
 			else
 				{
-				bn_wexpand(a,k);
-				bn_wexpand(b,k);
 				bn_wexpand(t,k*4);
 				bn_wexpand(rr,k*4);
-				for (i=a->top; i<k; i++)
-					a->d[i]=0;
-				for (i=b->top; i<k; i++)
-					b->d[i]=0;
 				bn_mul_part_recursive(rr->d,a->d,b->d,al-j,j,t->d);
 				}
 			rr->top=top;
 			goto end;
 			}
+#endif
 		}
 #endif /* BN_RECURSION */
 	if (bn_wexpand(rr,top) == NULL) goto err;
@@ -740,7 +1092,7 @@ void bn_mul_normal(BN_ULONG *r, BN_ULONG *a, int na, BN_ULONG *b, int nb)
 	BN_ULONG *rr;
 
 #ifdef BN_COUNT
-	printf(" bn_mul_normal %d * %d\n",na,nb);
+	fprintf(stderr," bn_mul_normal %d * %d\n",na,nb);
 #endif
 
 	if (na < nb)
@@ -774,7 +1126,7 @@ void bn_mul_normal(BN_ULONG *r, BN_ULONG *a, int na, BN_ULONG *b, int nb)
 void bn_mul_low_normal(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
 	{
 #ifdef BN_COUNT
-	printf(" bn_mul_low_normal %d * %d\n",n,n);
+	fprintf(stderr," bn_mul_low_normal %d * %d\n",n,n);
 #endif
 	bn_mul_words(r,a,n,b[0]);