Extra i386+gcc bn_div.c tune-up featuring inline division and saving
authorAndy Polyakov <appro@openssl.org>
Sat, 31 Jul 1999 23:27:41 +0000 (23:27 +0000)
committerAndy Polyakov <appro@openssl.org>
Sat, 31 Jul 1999 23:27:41 +0000 (23:27 +0000)
the remainder left in %edx. Here is the resulting performance improvement
matrix (improvement as a result of this *and* previous tune-up committed
two days ago). The results were obtained by profiling the "div" part of
the crypto/bn/bnspeed.c.

CPU BN_div bn_div_words overall comment
------------------------------------------------------------------------
PII +16% accumulated by +2-3% PII multiplies damn fast! Taking
inlining multiplication out of the loop
didn't make too much difference.
Eliminating of the multiplication
involved in remainder calculation
is the major factor.

Pentium +45% accumulated by +7-9% mull isn't that fast and replacing
inlining multiplications with additions in
the loop has more visible effect:-)

MIPS +75% +12% +20-25% In addition to the taking mults
R10000 out of the loop (giving 12% in the
asm/mips3.s) three mults were
eliminated in BN_div.

Alpha +30% +50% +10-15% Same as above. But remember that
EV4 bn_div_words is a C implementation.
It takes 4 Alpha mults in C to do
the same thing as 1 MIPS mult in
assembler does. So the effect (50%)
is more impressive. But not the
overall one... Well, if Alpha
bn_mul_add would be implemented
in assembler overall improvement
would be closer to MIPS...

crypto/bn/bn_div.c

index 03b9152241798e5154c619c3f1b5ea0d332cd995..e4253f6095af82118bf20978e102de5711c532f6 100644 (file)
@@ -204,15 +204,41 @@ int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor,
 #ifdef BN_DIV3W
                q=bn_div_3_words(wnump,d0,d1);
 #else
-               BN_ULONG n0,n1,rem;
+
+#if !defined(NO_ASM)
+# 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;
 
                n0=wnump[0];
                n1=wnump[-1];
                if (n0 == d0)
                        q=BN_MASK2;
                else
-#if defined(BN_LLONG) && defined(BN_DIV2W)
-                       q=((((BN_ULLONG)n0)<<BN_BITS2)|n1)/((BN_ULLONG)d0);
+#if defined(BN_LLONG) && defined(BN_DIV2W) && !defined(bn_div_words)
+                       q=((((BN_ULLONG)n0)<<BN_BITS2)|n1)/d0;
 #else
                        q=bn_div_words(n0,n1,d0);
 #endif
@@ -220,13 +246,15 @@ int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor,
 #ifdef BN_LLONG
                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 (;;)
                        {
                         if (t2 <= ((((BN_ULLONG)rem)<<BN_BITS2)|wnump[-2]))
@@ -239,12 +267,13 @@ int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor,
 #else
                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; */
@@ -261,7 +290,7 @@ int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor,
                        }
 #endif
                }
-#endif /* BN_DIV3W */
+#endif /* !BN_DIV3W */
                wnum.d--; wnum.top++;
                l0=bn_mul_words(tmp->d,sdiv->d,div_n,q);
                tmp->d[div_n]=l0;