1 files changed, 36 insertions, 18 deletions

diff --git a/src/lib/libcrypto/bn/asm/x86_64-gcc.c b/src/lib/libcrypto/bn/asm/x86_64-gcc.c
index 450e8e4322..7378344251 100644
--- a/src/lib/libcrypto/bn/asm/x86_64-gcc.c
+++ b/src/lib/libcrypto/bn/asm/x86_64-gcc.c
@@ -13,20 +13,42 @@
  * A. Well, that's because this code is basically a quick-n-dirty
  *    proof-of-concept hack. As you can see it's implemented with
  *    inline assembler, which means that you're bound to GCC and that
- *    there must be a room for fine-tuning.
+ *    there might be enough room for further improvement.
  *
  * Q. Why inline assembler?
- * A. x86_64 features own ABI I'm not familiar with. Which is why
- *    I decided to let the compiler take care of subroutine
- *    prologue/epilogue as well as register allocation.
+ * A. x86_64 features own ABI which I'm not familiar with. This is
+ *    why I decided to let the compiler take care of subroutine
+ *    prologue/epilogue as well as register allocation. For reference.
+ *    Win64 implements different ABI for AMD64, different from Linux.
  *
  * Q. How much faster does it get?
- * A. Unfortunately people sitting on x86_64 hardware are prohibited
- *    to disclose the performance numbers, so they (SuSE labs to be
- *    specific) wouldn't tell me. However! Very similar coding technique
- *    (reaching out for 128-bit result from 64x64-bit multiplication)
- *    results in >3 times performance improvement on MIPS and I see no
- *    reason why gain on x86_64 would be so much different:-)
+ * A. 'apps/openssl speed rsa dsa' output with no-asm:
+ *
+ *                        sign    verify    sign/s verify/s
+ *      rsa  512 bits   0.0006s   0.0001s   1683.8  18456.2
+ *      rsa 1024 bits   0.0028s   0.0002s    356.0   6407.0
+ *      rsa 2048 bits   0.0172s   0.0005s     58.0   1957.8
+ *      rsa 4096 bits   0.1155s   0.0018s      8.7    555.6
+ *                        sign    verify    sign/s verify/s
+ *      dsa  512 bits   0.0005s   0.0006s   2100.8   1768.3
+ *      dsa 1024 bits   0.0014s   0.0018s    692.3    559.2
+ *      dsa 2048 bits   0.0049s   0.0061s    204.7    165.0
+ *
+ *    'apps/openssl speed rsa dsa' output with this module:
+ *
+ *                        sign    verify    sign/s verify/s
+ *      rsa  512 bits   0.0004s   0.0000s   2767.1  33297.9
+ *      rsa 1024 bits   0.0012s   0.0001s    867.4  14674.7
+ *      rsa 2048 bits   0.0061s   0.0002s    164.0   5270.0
+ *      rsa 4096 bits   0.0384s   0.0006s     26.1   1650.8
+ *                        sign    verify    sign/s verify/s
+ *      dsa  512 bits   0.0002s   0.0003s   4442.2   3786.3
+ *      dsa 1024 bits   0.0005s   0.0007s   1835.1   1497.4
+ *      dsa 2048 bits   0.0016s   0.0020s    620.4    504.6
+ *
+ *    For the reference. IA-32 assembler implementation performs
+ *    very much like 64-bit code compiled with no-asm on the same
+ *    machine.
  */
 
 #define BN_ULONG unsigned long
@@ -151,7 +173,7 @@ BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d)
 }
 
 BN_ULONG bn_add_words (BN_ULONG *rp, BN_ULONG *ap, BN_ULONG *bp,int n)
-{ BN_ULONG ret,i;
+{ BN_ULONG ret=0,i=0;
 
         if (n <= 0) return 0;
 
@@ -164,7 +186,7 @@ BN_ULONG bn_add_words (BN_ULONG *rp, BN_ULONG *ap, BN_ULONG *bp,int n)
         "       leaq    1(%2),%2        \n"
         "       loop    1b              \n"
         "       sbbq    %0,%0           \n"
-                : "+a"(ret),"+c"(n),"+r"(i)
+                : "=&a"(ret),"+c"(n),"=&r"(i)
                 : "r"(rp),"r"(ap),"r"(bp)
                 : "cc"
         );
@@ -174,7 +196,7 @@ BN_ULONG bn_add_words (BN_ULONG *rp, BN_ULONG *ap, BN_ULONG *bp,int n)
 
 #ifndef SIMICS
 BN_ULONG bn_sub_words (BN_ULONG *rp, BN_ULONG *ap, BN_ULONG *bp,int n)
-{ BN_ULONG ret,i;
+{ BN_ULONG ret=0,i=0;
 
         if (n <= 0) return 0;
 
@@ -187,7 +209,7 @@ BN_ULONG bn_sub_words (BN_ULONG *rp, BN_ULONG *ap, BN_ULONG *bp,int n)
         "       leaq    1(%2),%2        \n"
         "       loop    1b              \n"
         "       sbbq    %0,%0           \n"
-                : "+a"(ret),"+c"(n),"+r"(i)
+                : "=&a"(ret),"+c"(n),"=&r"(i)
                 : "r"(rp),"r"(ap),"r"(bp)
                 : "cc"
         );
@@ -318,7 +340,6 @@ BN_ULONG bn_sub_words(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
 
 void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
         {
-        BN_ULONG bl,bh;
         BN_ULONG t1,t2;
         BN_ULONG c1,c2,c3;
 
@@ -423,7 +444,6 @@ void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
 
 void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
         {
-        BN_ULONG bl,bh;
         BN_ULONG t1,t2;
         BN_ULONG c1,c2,c3;
 
@@ -464,7 +484,6 @@ void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
 
 void bn_sqr_comba8(BN_ULONG *r, BN_ULONG *a)
         {
-        BN_ULONG bl,bh;
         BN_ULONG t1,t2;
         BN_ULONG c1,c2,c3;
 
@@ -541,7 +560,6 @@ void bn_sqr_comba8(BN_ULONG *r, BN_ULONG *a)
 
 void bn_sqr_comba4(BN_ULONG *r, BN_ULONG *a)
         {
-        BN_ULONG bl,bh;
         BN_ULONG t1,t2;
         BN_ULONG c1,c2,c3;