X-Git-Url: https://git.librecmc.org/?a=blobdiff_plain;f=crypto%2Fbn%2Fbn_div.c;h=830da87f58e047f7156a0b2aacd2195bb3495480;hb=f18a93ab04f248de45a8bcdded9b91880c690dbd;hp=defcf90c8223d282b36306aa3737dc3b22ca7460;hpb=adc7fe123bffef5514dd0fe18e184623a589b3d3;p=oweals%2Fopenssl.git

diff --git a/crypto/bn/bn_div.c b/crypto/bn/bn_div.c
index defcf90c82..830da87f58 100644
--- a/crypto/bn/bn_div.c
+++ b/crypto/bn/bn_div.c
@@ -200,57 +200,99 @@ int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor,
 
 	for (i=0; i<loop-1; i++)
 		{
-		BN_ULONG q,n0,n1;
-		BN_ULONG l0;
+		BN_ULONG q,l0;
+#ifdef BN_DIV3W
+		q=bn_div_3_words(wnump,d1,d0);
+#else
+
+#if !defined(NO_ASM) && !defined(PEDANTIC)
+# if defined(__GNUC__) && __GNUC__>=2
+#  if defined(__i386)
+   /*
+    * There were two reasons for implementing this template:
+    * - GNU C generates a call to a function (__udivdi3 to be exact)
+    *   in reply to ((((BN_ULLONG)n0)<<BN_BITS2)|n1)/d0 (I fail to
+    *   understand why...);
+    * - divl doesn't only calculate quotient, but also leaves
+    *   remainder in %edx which we can definitely use here:-)
+    *
+    *					<appro@fy.chalmers.se>
+    */
+#  define bn_div_words(n0,n1,d0)		\
+	({  asm volatile (			\
+		"divl	%4"			\
+		: "=a"(q), "=d"(rem)		\
+		: "a"(n1), "d"(n0), "g"(d0)	\
+		: "cc");			\
+	    q;					\
+	})
+#  define REMINDER_IS_ALREADY_CALCULATED
+#  endif /* __<cpu> */
+# endif /* __GNUC__ */
+#endif /* NO_ASM */
+		BN_ULONG n0,n1,rem=0;
 
-		wnum.d--; wnum.top++;
 		n0=wnump[0];
 		n1=wnump[-1];
 		if (n0 == d0)
 			q=BN_MASK2;
 		else
+#if defined(BN_LLONG) && defined(BN_DIV2W) && !defined(bn_div_words)
+			q=(BN_ULONG)(((((BN_ULLONG)n0)<<BN_BITS2)|n1)/d0);
+#else
 			q=bn_div_words(n0,n1,d0);
+#endif
 		{
 #ifdef BN_LLONG
-		BN_ULLONG t1,t2,rem;
-		t1=((BN_ULLONG)n0<<BN_BITS2)|n1;
+		BN_ULLONG t2;
+
+#ifndef REMINDER_IS_ALREADY_CALCULATED
+		/*
+		 * rem doesn't have to be BN_ULLONG. The least we
+		 * know it's less that d0, isn't it?
+		 */
+		rem=(n1-q*d0)&BN_MASK2;
+#endif
+		t2=(BN_ULLONG)d1*q;
+
 		for (;;)
 			{
-			rem=t1-(BN_ULLONG)q*d0;
-			t2=(BN_ULLONG)d1*q;
-                        if ((rem>>BN_BITS2) ||
-				(t2 <= ((rem<<BN_BITS2)|wnump[-2])))
+                        if (t2 <= ((((BN_ULLONG)rem)<<BN_BITS2)|wnump[-2]))
 				break;
 			q--;
+			rem += d0;
+			if (rem < d0) break; /* don't let rem overflow */
+			t2 -= d1;
 			}
 #else
-		BN_ULONG t1l,t1h,t2l,t2h,t3l,t3h,ql,qh,t3t;
-		t1h=n0;
-		t1l=n1;
+		BN_ULONG t2l,t2h,ql,qh;
+
+#ifndef REMINDER_IS_ALREADY_CALCULATED
+		/*
+		 * It's more than enough with the only multiplication.
+		 * See the comment above in BN_LLONG section...
+		 */
+		rem=(n1-q*d0)&BN_MASK2;
+#endif
+		t2l=LBITS(d1); t2h=HBITS(d1);
+		ql =LBITS(q);  qh =HBITS(q);
+		mul64(t2l,t2h,ql,qh); /* t2=(BN_ULLONG)d1*q; */
+
 		for (;;)
 			{
-			t2l=LBITS(d1); t2h=HBITS(d1);
-			ql =LBITS(q);  qh =HBITS(q);
-			mul64(t2l,t2h,ql,qh); /* t2=(BN_ULLONG)d1*q; */
-
-			t3t=LBITS(d0); t3h=HBITS(d0);
-			mul64(t3t,t3h,ql,qh); /* t3=t1-(BN_ULLONG)q*d0; */
-			t3l=(t1l-t3t)&BN_MASK2;
-			if (t3l > t1l) t3h++;
-			t3h=(t1h-t3h)&BN_MASK2;
-
-			/*if ((t3>>BN_BITS2) ||
-				(t2 <= ((t3<<BN_BITS2)+wnump[-2])))
-				break; */
-			if (t3h) break;
-			if (t2h < t3l) break;
-			if ((t2h == t3l) && (t2l <= wnump[-2])) break;
-
+			if ((t2h < rem) ||
+				((t2h == rem) && (t2l <= wnump[-2])))
+				break;
 			q--;
+			rem += d0;
+			if (rem < d0) break; /* don't let rem overflow */
+			if (t2l < d1) t2h--; t2l -= d1;
 			}
 #endif
 		}
+#endif /* !BN_DIV3W */
 		l0=bn_mul_words(tmp->d,sdiv->d,div_n,q);
+		wnum.d--; wnum.top++;
 		tmp->d[div_n]=l0;
 		for (j=div_n+1; j>0; j--)
 			if (tmp->d[j-1]) break;