import openssl-0.9.7-beta1

10 files changed, 4545 insertions, 1296 deletions

diff --git a/src/lib/libcrypto/bn/asm/README b/src/lib/libcrypto/bn/asm/README
index d93fbff77f..b0f3a68a06 100644
--- a/src/lib/libcrypto/bn/asm/README
+++ b/src/lib/libcrypto/bn/asm/README
@@ -1,5 +1,7 @@
+<OBSOLETE>
+
 All assember in this directory are just version of the file
-crypto/bn/bn_mulw.c.
+crypto/bn/bn_asm.c.
 
 Quite a few of these files are just the assember output from gcc since on 
 quite a few machines they are 2 times faster than the system compiler.
@@ -15,16 +17,11 @@ On the 2 alpha C compilers I had access to, it was not possible to do
 were 64 bits).  So the hand assember gives access to the 128 bit result and
 a 2 times speedup :-).
 
-The x86xxxx.obj files are the assembled version of x86xxxx.asm files.
-I had such a hard time finding a macro assember for Microsoft, I decided to
-include the object file to save others the hassle :-).
+There are 3 versions of assember for the HP PA-RISC.
+
+pa-risc.s is the origional one which works fine and generated using gcc :-)
 
-I have also included uu encoded versions of the .obj incase they get
-trashed.
+pa-risc2W.s and pa-risc2.s are 64 and 32-bit PA-RISC 2.0 implementations
+by Chris Ruemmler from HP (with some help from the HP C compiler).
 
-There are 2 versions of assember for the HP PA-RISC.
-pa-risc.s is the origional one which works fine.
-pa-risc2.s is a new version that often generates warnings but if the
-tests pass, it gives performance that is over 2 times faster than
-pa-risc.s.
-Both were generated using gcc :-)
+</OBSOLETE>
diff --git a/src/lib/libcrypto/bn/asm/alpha.s b/src/lib/libcrypto/bn/asm/alpha.s
index 1d17b1d619..555ff0b92d 100644
--- a/src/lib/libcrypto/bn/asm/alpha.s
+++ b/src/lib/libcrypto/bn/asm/alpha.s
@@ -1,8 +1,14 @@
  # DEC Alpha assember
- # The bn_div64 is actually gcc output but the other parts are hand done.
+ # The bn_div_words is actually gcc output but the other parts are hand done.
  # Thanks to tzeruch@ceddec.com for sending me the gcc output for
- # bn_div64.
-        .file   1 "bn_mulw.c"
+ # bn_div_words.
+ # I've gone back and re-done most of routines.
+ # The key thing to remeber for the 164 CPU is that while a
+ # multiply operation takes 8 cycles, another one can only be issued
+ # after 4 cycles have elapsed.  I've done modification to help
+ # improve this.  Also, normally, a ld instruction will not be available
+ # for about 3 cycles.
+        .file   1 "bn_asm.c"
         .set noat
 gcc2_compiled.:
 __gnu_compiled_c:
@@ -14,65 +20,91 @@ bn_mul_add_words:
 bn_mul_add_words..ng:
         .frame $30,0,$26,0
         .prologue 0
-        subq $18,2,$25  # num=-2
-        bis $31,$31,$0
-        blt $25,$42
         .align 5
-$142:
-        subq $18,2,$18  # num-=2
-        subq $25,2,$25  # num-=2
-
-        ldq $1,0($17)   # a[0]
-        ldq $2,8($17)   # a[1]
-
-        mulq $19,$1,$3  # a[0]*w low part       r3
-        umulh $19,$1,$1 # a[0]*w high part      r1
-        mulq $19,$2,$4  # a[1]*w low part       r4
-        umulh $19,$2,$2 # a[1]*w high part      r2
-
-        ldq $22,0($16)  # r[0]                  r22
-        ldq $23,8($16)  # r[1]                  r23
-
-        addq $3,$22,$3  # a0 low part + r[0]    
-        addq $4,$23,$4  # a1 low part + r[1]    
-        cmpult $3,$22,$5 # overflow?
-        cmpult $4,$23,$6 # overflow?
-        addq $5,$1,$1   # high part + overflow 
-        addq $6,$2,$2   # high part + overflow 
-
-        addq $3,$0,$3   # add c
-        cmpult $3,$0,$5 # overflow?
-        stq $3,0($16)
-        addq $5,$1,$0   # c=high part + overflow 
-
-        addq $4,$0,$4   # add c
-        cmpult $4,$0,$5 # overflow?
-        stq $4,8($16)
-        addq $5,$2,$0   # c=high part + overflow 
+        subq    $18,4,$18
+        bis     $31,$31,$0
+        blt     $18,$43         # if we are -1, -2, -3 or -4 goto tail code
+        ldq     $20,0($17)      # 1 1
+        ldq     $1,0($16)       # 1 1
+        .align 3
+$42:
+        mulq    $20,$19,$5      # 1 2 1 ######
+        ldq     $21,8($17)      # 2 1
+        ldq     $2,8($16)       # 2 1
+        umulh   $20,$19,$20     # 1 2   ######
+        ldq     $27,16($17)     # 3 1
+        ldq     $3,16($16)      # 3 1
+        mulq    $21,$19,$6      # 2 2 1 ######
+         ldq    $28,24($17)     # 4 1
+        addq    $1,$5,$1        # 1 2 2
+         ldq    $4,24($16)      # 4 1
+        umulh   $21,$19,$21     # 2 2   ######
+         cmpult $1,$5,$22       # 1 2 3 1
+        addq    $20,$22,$20     # 1 3 1
+         addq   $1,$0,$1        # 1 2 3 1
+        mulq    $27,$19,$7      # 3 2 1 ######
+         cmpult $1,$0,$0        # 1 2 3 2
+        addq    $2,$6,$2        # 2 2 2
+         addq   $20,$0,$0       # 1 3 2 
+        cmpult  $2,$6,$23       # 2 2 3 1
+         addq   $21,$23,$21     # 2 3 1
+        umulh   $27,$19,$27     # 3 2   ######
+         addq   $2,$0,$2        # 2 2 3 1
+        cmpult  $2,$0,$0        # 2 2 3 2
+         subq   $18,4,$18
+        mulq    $28,$19,$8      # 4 2 1 ######
+         addq   $21,$0,$0       # 2 3 2 
+        addq    $3,$7,$3        # 3 2 2
+         addq   $16,32,$16
+        cmpult  $3,$7,$24       # 3 2 3 1
+         stq    $1,-32($16)     # 1 2 4
+        umulh   $28,$19,$28     # 4 2   ######
+         addq   $27,$24,$27     # 3 3 1
+        addq    $3,$0,$3        # 3 2 3 1
+         stq    $2,-24($16)     # 2 2 4
+        cmpult  $3,$0,$0        # 3 2 3 2
+         stq    $3,-16($16)     # 3 2 4
+        addq    $4,$8,$4        # 4 2 2
+         addq   $27,$0,$0       # 3 3 2 
+        cmpult  $4,$8,$25       # 4 2 3 1
+         addq   $17,32,$17
+        addq    $28,$25,$28     # 4 3 1
+         addq   $4,$0,$4        # 4 2 3 1
+        cmpult  $4,$0,$0        # 4 2 3 2
+         stq    $4,-8($16)      # 4 2 4
+        addq    $28,$0,$0       # 4 3 2 
+         blt    $18,$43
 
-        ble $18,$43
+        ldq     $20,0($17)      # 1 1
+        ldq     $1,0($16)       # 1 1
 
-        addq $16,16,$16
-        addq $17,16,$17
-        blt $25,$42
+        br      $42
 
-        br $31,$142
-$42:
-        ldq $1,0($17)   # a[0]
-        umulh $19,$1,$3 # a[0]*w high part
-        mulq $19,$1,$1  # a[0]*w low part
-        ldq $2,0($16)   # r[0]
-        addq $1,$2,$1   # low part + r[0]
-        cmpult $1,$2,$4 # overflow?
-        addq $4,$3,$3   # high part + overflow 
-        addq $1,$0,$1   # add c
-        cmpult $1,$0,$4 # overflow?
-        addq $4,$3,$0   # c=high part + overflow 
-        stq $1,0($16)
+        .align 4
+$45:
+        ldq     $20,0($17)      # 4 1
+        ldq     $1,0($16)       # 4 1
+        mulq    $20,$19,$5      # 4 2 1
+        subq    $18,1,$18
+        addq    $16,8,$16
+        addq    $17,8,$17
+        umulh   $20,$19,$20     # 4 2
+        addq    $1,$5,$1        # 4 2 2
+        cmpult  $1,$5,$22       # 4 2 3 1
+        addq    $20,$22,$20     # 4 3 1
+        addq    $1,$0,$1        # 4 2 3 1
+        cmpult  $1,$0,$0        # 4 2 3 2
+        addq    $20,$0,$0       # 4 3 2 
+        stq     $1,-8($16)      # 4 2 4
+        bgt     $18,$45
+        ret     $31,($26),1     # else exit
 
         .align 4
 $43:
-        ret $31,($26),1
+        addq    $18,4,$18
+        bgt     $18,$45         # goto tail code
+        ret     $31,($26),1     # else exit
+
         .end bn_mul_add_words
         .align 3
         .globl bn_mul_words
@@ -81,49 +113,75 @@ bn_mul_words:
 bn_mul_words..ng:
         .frame $30,0,$26,0
         .prologue 0
-        subq $18,2,$25  # num=-2
-        bis $31,$31,$0
-        blt $25,$242
         .align 5
-$342:
-        subq $18,2,$18  # num-=2
-        subq $25,2,$25  # num-=2
-
-        ldq $1,0($17)   # a[0]
-        ldq $2,8($17)   # a[1]
-
-        mulq $19,$1,$3  # a[0]*w low part       r3
-        umulh $19,$1,$1 # a[0]*w high part      r1
-        mulq $19,$2,$4  # a[1]*w low part       r4
-        umulh $19,$2,$2 # a[1]*w high part      r2
-
-        addq $3,$0,$3   # add c
-        cmpult $3,$0,$5 # overflow?
-        stq $3,0($16)
-        addq $5,$1,$0   # c=high part + overflow 
-
-        addq $4,$0,$4   # add c
-        cmpult $4,$0,$5 # overflow?
-        stq $4,8($16)
-        addq $5,$2,$0   # c=high part + overflow 
-
-        ble $18,$243
-
-        addq $16,16,$16
-        addq $17,16,$17
-        blt $25,$242
-
-        br $31,$342
-$242:
-        ldq $1,0($17)   # a[0]
-        umulh $19,$1,$3 # a[0]*w high part
-        mulq $19,$1,$1  # a[0]*w low part
-        addq $1,$0,$1   # add c
-        cmpult $1,$0,$4 # overflow?
-        addq $4,$3,$0   # c=high part + overflow 
-        stq $1,0($16)
-$243:
-        ret $31,($26),1
+        subq    $18,4,$18
+        bis     $31,$31,$0
+        blt     $18,$143        # if we are -1, -2, -3 or -4 goto tail code
+        ldq     $20,0($17)      # 1 1
+        .align 3
+$142:
+
+        mulq    $20,$19,$5      # 1 2 1 #####
+         ldq    $21,8($17)      # 2 1
+         ldq    $27,16($17)     # 3 1
+        umulh   $20,$19,$20     # 1 2   #####
+         ldq    $28,24($17)     # 4 1
+        mulq    $21,$19,$6      # 2 2 1 #####
+         addq   $5,$0,$5        # 1 2 3 1
+        subq    $18,4,$18
+         cmpult $5,$0,$0        # 1 2 3 2
+        umulh   $21,$19,$21     # 2 2   #####
+         addq   $20,$0,$0       # 1 3 2 
+        addq    $17,32,$17
+         addq   $6,$0,$6        # 2 2 3 1
+        mulq    $27,$19,$7      # 3 2 1 #####
+         cmpult $6,$0,$0        # 2 2 3 2
+        addq    $21,$0,$0       # 2 3 2 
+         addq   $16,32,$16
+        umulh   $27,$19,$27     # 3 2   #####
+         stq    $5,-32($16)     # 1 2 4
+        mulq    $28,$19,$8      # 4 2 1 #####
+         addq   $7,$0,$7        # 3 2 3 1
+        stq     $6,-24($16)     # 2 2 4
+         cmpult $7,$0,$0        # 3 2 3 2
+        umulh   $28,$19,$28     # 4 2   #####
+         addq   $27,$0,$0       # 3 3 2 
+        stq     $7,-16($16)     # 3 2 4
+         addq   $8,$0,$8        # 4 2 3 1
+        cmpult  $8,$0,$0        # 4 2 3 2
+
+        addq    $28,$0,$0       # 4 3 2 
+
+        stq     $8,-8($16)      # 4 2 4
+
+        blt     $18,$143
+
+        ldq     $20,0($17)      # 1 1
+
+        br      $142
+
+        .align 4
+$145:
+        ldq     $20,0($17)      # 4 1
+        mulq    $20,$19,$5      # 4 2 1
+        subq    $18,1,$18
+        umulh   $20,$19,$20     # 4 2
+        addq    $5,$0,$5        # 4 2 3 1
+         addq   $16,8,$16
+        cmpult  $5,$0,$0        # 4 2 3 2
+         addq   $17,8,$17
+        addq    $20,$0,$0       # 4 3 2 
+        stq     $5,-8($16)      # 4 2 4
+
+        bgt     $18,$145
+        ret     $31,($26),1     # else exit
+
+        .align 4
+$143:
+        addq    $18,4,$18
+        bgt     $18,$145        # goto tail code
+        ret     $31,($26),1     # else exit
+
         .end bn_mul_words
         .align 3
         .globl bn_sqr_words
@@ -132,44 +190,58 @@ bn_sqr_words:
 bn_sqr_words..ng:
         .frame $30,0,$26,0
         .prologue 0
-        
-        subq $18,2,$25  # num=-2
-        blt $25,$442
-        .align 5
-$542:
-        subq $18,2,$18  # num-=2
-        subq $25,2,$25  # num-=2
 
-        ldq $1,0($17)   # a[0]
-        ldq $4,8($17)   # a[1]
-
-        mulq $1,$1,$2   # a[0]*w low part       r2
-        umulh $1,$1,$3 # a[0]*w high part       r3
-        mulq $4,$4,$5   # a[1]*w low part       r5
-        umulh $4,$4,$6 # a[1]*w high part       r6
-
-        stq $2,0($16)   # r[0]
-        stq $3,8($16)   # r[1]
-        stq $5,16($16)  # r[3]
-        stq $6,24($16)  # r[4]
+        subq    $18,4,$18
+        blt     $18,$543        # if we are -1, -2, -3 or -4 goto tail code
+        ldq     $20,0($17)      # 1 1
+        .align 3
+$542:
+        mulq    $20,$20,$5              ######
+         ldq    $21,8($17)      # 1 1
+        subq    $18,4
+        umulh   $20,$20,$1              ######
+        ldq     $27,16($17)     # 1 1
+        mulq    $21,$21,$6              ######
+        ldq     $28,24($17)     # 1 1
+        stq     $5,0($16)       # r[0]
+        umulh   $21,$21,$2              ######
+        stq     $1,8($16)       # r[1]
+        mulq    $27,$27,$7              ######
+        stq     $6,16($16)      # r[0]
+        umulh   $27,$27,$3              ######
+        stq     $2,24($16)      # r[1]
+        mulq    $28,$28,$8              ######
+        stq     $7,32($16)      # r[0]
+        umulh   $28,$28,$4              ######
+        stq     $3,40($16)      # r[1]
 
-        ble $18,$443
+        addq    $16,64,$16
+        addq    $17,32,$17
+        stq     $8,-16($16)     # r[0]
+        stq     $4,-8($16)      # r[1]
 
-        addq $16,32,$16
-        addq $17,16,$17
-        blt $25,$442
-        br $31,$542
+        blt     $18,$543
+        ldq     $20,0($17)      # 1 1
+        br      $542
 
 $442:
-        ldq $1,0($17)   # a[0]
-        mulq $1,$1,$2   # a[0]*w low part       r2
-        umulh $1,$1,$3  # a[0]*w high part       r3
-        stq $2,0($16)   # r[0]
-        stq $3,8($16)   # r[1]
+        ldq     $20,0($17)   # a[0]
+        mulq    $20,$20,$5  # a[0]*w low part       r2
+        addq    $16,16,$16
+        addq    $17,8,$17
+        subq    $18,1,$18
+        umulh   $20,$20,$1  # a[0]*w high part       r3
+        stq     $5,-16($16)   # r[0]
+        stq     $1,-8($16)   # r[1]
+
+        bgt     $18,$442
+        ret     $31,($26),1     # else exit
 
         .align 4
-$443:
-        ret $31,($26),1
+$543:
+        addq    $18,4,$18
+        bgt     $18,$442        # goto tail code
+        ret     $31,($26),1     # else exit
         .end bn_sqr_words
 
         .align 3
@@ -180,31 +252,74 @@ bn_add_words..ng:
         .frame $30,0,$26,0
         .prologue 0
 
-        bis     $31,$31,$8      # carry = 0
-        ble     $19,$900
+        subq    $19,4,$19
+        bis     $31,$31,$0      # carry = 0
+        blt     $19,$900
+        ldq     $5,0($17)       # a[0]
+        ldq     $1,0($18)       # b[1]
+        .align 3
 $901:
-        ldq     $0,0($17)       # a[0]
-        ldq     $1,0($18)       # a[1]
+        addq    $1,$5,$1        # r=a+b;
+         ldq    $6,8($17)       # a[1]
+        cmpult  $1,$5,$22       # did we overflow?
+         ldq    $2,8($18)       # b[1]
+        addq    $1,$0,$1        # c+= overflow
+         ldq    $7,16($17)      # a[2]
+        cmpult  $1,$0,$0        # overflow?
+         ldq    $3,16($18)      # b[2]
+        addq    $0,$22,$0
+         ldq    $8,24($17)      # a[3]
+        addq    $2,$6,$2        # r=a+b;
+         ldq    $4,24($18)      # b[3]
+        cmpult  $2,$6,$23       # did we overflow?
+         addq   $3,$7,$3        # r=a+b;
+        addq    $2,$0,$2        # c+= overflow
+         cmpult $3,$7,$24       # did we overflow?
+        cmpult  $2,$0,$0        # overflow?
+         addq   $4,$8,$4        # r=a+b;
+        addq    $0,$23,$0
+         cmpult $4,$8,$25       # did we overflow?
+        addq    $3,$0,$3        # c+= overflow
+         stq    $1,0($16)       # r[0]=c
+        cmpult  $3,$0,$0        # overflow?
+         stq    $2,8($16)       # r[1]=c
+        addq    $0,$24,$0
+         stq    $3,16($16)      # r[2]=c
+        addq    $4,$0,$4        # c+= overflow
+         subq   $19,4,$19       # loop--
+        cmpult  $4,$0,$0        # overflow?
+         addq   $17,32,$17      # a++
+        addq    $0,$25,$0
+         stq    $4,24($16)      # r[3]=c
+        addq    $18,32,$18      # b++
+         addq   $16,32,$16      # r++
 
-        addq    $0,$1,$3        # c=a+b;
+        blt     $19,$900
+         ldq    $5,0($17)       # a[0]
+        ldq     $1,0($18)       # b[1]
+         br     $901
+        .align 4
+$945:
+        ldq     $5,0($17)       # a[0]
+         ldq    $1,0($18)       # b[1]
+        addq    $1,$5,$1        # r=a+b;
+         subq   $19,1,$19       # loop--
+        addq    $1,$0,$1        # c+= overflow
          addq   $17,8,$17       # a++
+        cmpult  $1,$5,$22       # did we overflow?
+         cmpult $1,$0,$0        # overflow?
+        addq    $18,8,$18       # b++
+         stq    $1,0($16)       # r[0]=c
+        addq    $0,$22,$0
+         addq   $16,8,$16       # r++
 
-        cmpult  $3,$1,$7        # did we overflow?
-         addq   $18,8,$18       # b++
-
-        addq    $8,$3,$3        # c+=carry
-
-        cmpult  $3,$8,$8        # did we overflow?
-         stq    $3,($16)        # r[0]=c
-
-        addq    $7,$8,$8        # add into overflow
-         subq   $19,1,$19       # loop--
+        bgt     $19,$945
+        ret     $31,($26),1     # else exit
 
-        addq    $16,8,$16       # r++
-         bgt    $19,$901
 $900:
-        bis     $8,$8,$0        # return carry
-        ret $31,($26),1
+        addq    $19,4,$19
+        bgt     $19,$945        # goto tail code
+        ret     $31,($26),1     # else exit
         .end bn_add_words
 
  #
@@ -213,11 +328,11 @@ $900:
  #
 .text
         .align 3
-        .globl bn_div64
-        .ent bn_div64
-bn_div64:
+        .globl bn_div_words
+        .ent bn_div_words
+bn_div_words:
         ldgp $29,0($27)
-bn_div64..ng:
+bn_div_words..ng:
         lda $30,-48($30)
         .frame $30,48,$26,0
         stq $26,0($30)
@@ -338,7 +453,2747 @@ $136:
         ldq $13,40($30)
         addq $30,48,$30
         ret $31,($26),1
-        .end bn_div64
-        .ident  "GCC: (GNU) 2.7.2.1"
+        .end bn_div_words
+
+        .set noat
+        .text
+        .align 3
+        .globl bn_sub_words
+        .ent bn_sub_words
+bn_sub_words:
+bn_sub_words..ng:
+        .frame $30,0,$26,0
+        .prologue 0
+
+        subq    $19,    4,      $19
+        bis     $31,    $31,    $0
+        blt     $19,    $100
+        ldq     $1,     0($17)
+        ldq     $2,     0($18)
+$101:
+        ldq     $3,     8($17)
+        cmpult  $1,     $2,     $4
+        ldq     $5,     8($18)
+        subq    $1,     $2,     $1
+        ldq     $6,     16($17)
+        cmpult  $1,     $0,     $2
+        ldq     $7,     16($18)
+        subq    $1,     $0,     $23
+        ldq     $8,     24($17)
+        addq    $2,     $4,     $0
+        cmpult  $3,     $5,     $24
+        subq    $3,     $5,     $3
+        ldq     $22,    24($18)
+        cmpult  $3,     $0,     $5
+        subq    $3,     $0,     $25
+        addq    $5,     $24,    $0
+        cmpult  $6,     $7,     $27
+        subq    $6,     $7,     $6
+        stq     $23,    0($16)
+        cmpult  $6,     $0,     $7
+        subq    $6,     $0,     $28
+        addq    $7,     $27,    $0
+        cmpult  $8,     $22,    $21
+        subq    $8,     $22,    $8
+        stq     $25,    8($16)
+        cmpult  $8,     $0,     $22
+        subq    $8,     $0,     $20
+        addq    $22,    $21,    $0
+        stq     $28,    16($16)
+        subq    $19,    4,      $19
+        stq     $20,    24($16)
+        addq    $17,    32,     $17
+        addq    $18,    32,     $18
+        addq    $16,    32,     $16
+        blt     $19,    $100
+        ldq     $1,     0($17)
+        ldq     $2,     0($18)
+        br      $101
+$102:
+        ldq     $1,     0($17)
+        ldq     $2,     0($18)
+        cmpult  $1,     $2,     $27
+        subq    $1,     $2,     $1
+        cmpult  $1,     $0,     $2
+        subq    $1,     $0,     $1
+        stq     $1,     0($16)
+        addq    $2,     $27,    $0
+        addq    $17,    8,      $17
+        addq    $18,    8,      $18
+        addq    $16,    8,      $16
+        subq    $19,    1,      $19
+        bgt     $19,    $102
+        ret     $31,($26),1
+$100:
+        addq    $19,    4,      $19
+        bgt     $19,    $102
+$103:
+        ret     $31,($26),1
+        .end bn_sub_words
+        .text
+        .align 3
+        .globl bn_mul_comba4
+        .ent bn_mul_comba4
+bn_mul_comba4:
+bn_mul_comba4..ng:
+        .frame $30,0,$26,0
+        .prologue 0
 
+        ldq     $0,     0($17)
+        ldq     $1,     0($18)
+        ldq     $2,     8($17)
+        ldq     $3,     8($18)
+        ldq     $4,     16($17)
+        ldq     $5,     16($18)
+        ldq     $6,     24($17)
+        ldq     $7,     24($18)
+        bis     $31,    $31,    $23
+        mulq    $0,     $1,     $8
+        umulh   $0,     $1,     $22
+        stq     $8,     0($16)
+        bis     $31,    $31,    $8
+        mulq    $0,     $3,     $24
+        umulh   $0,     $3,     $25
+        addq    $22,    $24,    $22
+        cmpult  $22,    $24,    $27
+        addq    $27,    $25,    $25
+        addq    $23,    $25,    $23
+        cmpult  $23,    $25,    $28
+        addq    $8,     $28,    $8
+        mulq    $2,     $1,     $21
+        umulh   $2,     $1,     $20
+        addq    $22,    $21,    $22
+        cmpult  $22,    $21,    $19
+        addq    $19,    $20,    $20
+        addq    $23,    $20,    $23
+        cmpult  $23,    $20,    $17
+        addq    $8,     $17,    $8
+        stq     $22,    8($16)
+        bis     $31,    $31,    $22
+        mulq    $2,     $3,     $18
+        umulh   $2,     $3,     $24
+        addq    $23,    $18,    $23
+        cmpult  $23,    $18,    $27
+        addq    $27,    $24,    $24
+        addq    $8,     $24,    $8
+        cmpult  $8,     $24,    $25
+        addq    $22,    $25,    $22
+        mulq    $0,     $5,     $28
+        umulh   $0,     $5,     $21
+        addq    $23,    $28,    $23
+        cmpult  $23,    $28,    $19
+        addq    $19,    $21,    $21
+        addq    $8,     $21,    $8
+        cmpult  $8,     $21,    $20
+        addq    $22,    $20,    $22
+        mulq    $4,     $1,     $17
+        umulh   $4,     $1,     $18
+        addq    $23,    $17,    $23
+        cmpult  $23,    $17,    $27
+        addq    $27,    $18,    $18
+        addq    $8,     $18,    $8
+        cmpult  $8,     $18,    $24
+        addq    $22,    $24,    $22
+        stq     $23,    16($16)
+        bis     $31,    $31,    $23
+        mulq    $0,     $7,     $25
+        umulh   $0,     $7,     $28
+        addq    $8,     $25,    $8
+        cmpult  $8,     $25,    $19
+        addq    $19,    $28,    $28
+        addq    $22,    $28,    $22
+        cmpult  $22,    $28,    $21
+        addq    $23,    $21,    $23
+        mulq    $2,     $5,     $20
+        umulh   $2,     $5,     $17
+        addq    $8,     $20,    $8
+        cmpult  $8,     $20,    $27
+        addq    $27,    $17,    $17
+        addq    $22,    $17,    $22
+        cmpult  $22,    $17,    $18
+        addq    $23,    $18,    $23
+        mulq    $4,     $3,     $24
+        umulh   $4,     $3,     $25
+        addq    $8,     $24,    $8
+        cmpult  $8,     $24,    $19
+        addq    $19,    $25,    $25
+        addq    $22,    $25,    $22
+        cmpult  $22,    $25,    $28
+        addq    $23,    $28,    $23
+        mulq    $6,     $1,     $21
+        umulh   $6,     $1,     $0
+        addq    $8,     $21,    $8
+        cmpult  $8,     $21,    $20
+        addq    $20,    $0,     $0
+        addq    $22,    $0,     $22
+        cmpult  $22,    $0,     $27
+        addq    $23,    $27,    $23
+        stq     $8,     24($16)
+        bis     $31,    $31,    $8
+        mulq    $2,     $7,     $17
+        umulh   $2,     $7,     $18
+        addq    $22,    $17,    $22
+        cmpult  $22,    $17,    $24
+        addq    $24,    $18,    $18
+        addq    $23,    $18,    $23
+        cmpult  $23,    $18,    $19
+        addq    $8,     $19,    $8
+        mulq    $4,     $5,     $25
+        umulh   $4,     $5,     $28
+        addq    $22,    $25,    $22
+        cmpult  $22,    $25,    $21
+        addq    $21,    $28,    $28
+        addq    $23,    $28,    $23
+        cmpult  $23,    $28,    $20
+        addq    $8,     $20,    $8
+        mulq    $6,     $3,     $0
+        umulh   $6,     $3,     $27
+        addq    $22,    $0,     $22
+        cmpult  $22,    $0,     $1
+        addq    $1,     $27,    $27
+        addq    $23,    $27,    $23
+        cmpult  $23,    $27,    $17
+        addq    $8,     $17,    $8
+        stq     $22,    32($16)
+        bis     $31,    $31,    $22
+        mulq    $4,     $7,     $24
+        umulh   $4,     $7,     $18
+        addq    $23,    $24,    $23
+        cmpult  $23,    $24,    $19
+        addq    $19,    $18,    $18
+        addq    $8,     $18,    $8
+        cmpult  $8,     $18,    $2
+        addq    $22,    $2,     $22
+        mulq    $6,     $5,     $25
+        umulh   $6,     $5,     $21
+        addq    $23,    $25,    $23
+        cmpult  $23,    $25,    $28
+        addq    $28,    $21,    $21
+        addq    $8,     $21,    $8
+        cmpult  $8,     $21,    $20
+        addq    $22,    $20,    $22
+        stq     $23,    40($16)
+        bis     $31,    $31,    $23
+        mulq    $6,     $7,     $0
+        umulh   $6,     $7,     $1
+        addq    $8,     $0,     $8
+        cmpult  $8,     $0,     $27
+        addq    $27,    $1,     $1
+        addq    $22,    $1,     $22
+        cmpult  $22,    $1,     $17
+        addq    $23,    $17,    $23
+        stq     $8,     48($16)
+        stq     $22,    56($16)
+        ret     $31,($26),1
+        .end bn_mul_comba4
+        .text
+        .align 3
+        .globl bn_mul_comba8
+        .ent bn_mul_comba8
+bn_mul_comba8:
+bn_mul_comba8..ng:
+        .frame $30,0,$26,0
+        .prologue 0
+        ldq     $1,     0($17)
+        ldq     $2,     0($18)
+        zapnot  $1,     15,     $7
+        srl     $2,     32,     $8
+        mulq    $8,     $7,     $22
+        srl     $1,     32,     $6
+        zapnot  $2,     15,     $5
+        mulq    $5,     $6,     $4
+        mulq    $7,     $5,     $24
+        addq    $22,    $4,     $22
+        cmpult  $22,    $4,     $1
+        mulq    $6,     $8,     $3
+        beq     $1,     $173
+        bis     $31,    1,      $1
+        sll     $1,     32,     $1
+        addq    $3,     $1,     $3
+$173:
+        sll     $22,    32,     $4
+        addq    $24,    $4,     $24
+        stq     $24,    0($16)
+        ldq     $2,     0($17)
+        ldq     $1,     8($18)
+        zapnot  $2,     15,     $7
+        srl     $1,     32,     $8
+        mulq    $8,     $7,     $25
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $0
+        srl     $2,     32,     $6
+        mulq    $5,     $6,     $23
+        mulq    $6,     $8,     $6
+        srl     $22,    32,     $1
+        cmpult  $24,    $4,     $2
+        addq    $3,     $1,     $3
+        addq    $2,     $3,     $22
+        addq    $25,    $23,    $25
+        cmpult  $25,    $23,    $1
+        bis     $31,    1,      $2
+        beq     $1,     $177
+        sll     $2,     32,     $1
+        addq    $6,     $1,     $6
+$177:
+        sll     $25,    32,     $23
+        ldq     $1,     0($18)
+        addq    $0,     $23,    $0
+        bis     $0,     $0,     $7
+        ldq     $3,     8($17)
+        addq    $22,    $7,     $22
+        srl     $1,     32,     $8
+        cmpult  $22,    $7,     $4
+        zapnot  $3,     15,     $7
+        mulq    $8,     $7,     $28
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $21
+        srl     $25,    32,     $1
+        cmpult  $0,     $23,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $4,     $6,     $24
+        srl     $3,     32,     $6
+        mulq    $5,     $6,     $2
+        mulq    $6,     $8,     $6
+        addq    $28,    $2,     $28
+        cmpult  $28,    $2,     $1
+        bis     $31,    1,      $2
+        beq     $1,     $181
+        sll     $2,     32,     $1
+        addq    $6,     $1,     $6
+$181:
+        sll     $28,    32,     $2
+        addq    $21,    $2,     $21
+        bis     $21,    $21,    $7
+        addq    $22,    $7,     $22
+        stq     $22,    8($16)
+        ldq     $3,     16($17)
+        ldq     $1,     0($18)
+        cmpult  $22,    $7,     $4
+        zapnot  $3,     15,     $7
+        srl     $1,     32,     $8
+        mulq    $8,     $7,     $22
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $20
+        srl     $28,    32,     $1
+        cmpult  $21,    $2,     $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $4,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $23
+        srl     $3,     32,     $6
+        mulq    $5,     $6,     $2
+        mulq    $6,     $8,     $6
+        addq    $22,    $2,     $22
+        cmpult  $22,    $2,     $1
+        bis     $31,    1,      $2
+        beq     $1,     $185
+        sll     $2,     32,     $1
+        addq    $6,     $1,     $6
+$185:
+        sll     $22,    32,     $2
+        ldq     $1,     8($18)
+        addq    $20,    $2,     $20
+        bis     $20,    $20,    $7
+        ldq     $4,     8($17)
+        addq    $24,    $7,     $24
+        srl     $1,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $25
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $0
+        srl     $22,    32,     $1
+        cmpult  $20,    $2,     $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $22
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $5
+        bis     $31,    1,      $21
+        addq    $25,    $5,     $25
+        cmpult  $25,    $5,     $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $189
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$189:
+        sll     $25,    32,     $5
+        ldq     $2,     16($18)
+        addq    $0,     $5,     $0
+        bis     $0,     $0,     $7
+        ldq     $4,     0($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $25,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $0,     $5,     $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $1,     $22,    $22
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $193
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$193:
+        sll     $28,    32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $24,    $7,     $24
+        stq     $24,    16($16)
+        ldq     $4,     0($17)
+        ldq     $5,     24($18)
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $24
+        mulq    $7,     $5,     $2
+        addq    $1,     $22,    $22
+        addq    $0,     $24,    $0
+        cmpult  $0,     $24,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $197
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$197:
+        sll     $0,     32,     $24
+        ldq     $1,     16($18)
+        addq    $2,     $24,    $2
+        bis     $2,     $2,     $7
+        ldq     $4,     8($17)
+        addq    $23,    $7,     $23
+        srl     $1,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $25
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $21
+        srl     $0,     32,     $1
+        cmpult  $2,     $24,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $24
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $5
+        bis     $31,    1,      $20
+        addq    $25,    $5,     $25
+        cmpult  $25,    $5,     $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $201
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$201:
+        sll     $25,    32,     $5
+        ldq     $2,     8($18)
+        addq    $21,    $5,     $21
+        bis     $21,    $21,    $7
+        ldq     $4,     16($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $25,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $21,    $5,     $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $24,    $24
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $205
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$205:
+        sll     $28,    32,     $25
+        ldq     $2,     0($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     24($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $1,     $24,    $24
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $209
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$209:
+        sll     $0,     32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $23,    $7,     $23
+        stq     $23,    24($16)
+        ldq     $4,     32($17)
+        ldq     $5,     0($18)
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $23
+        mulq    $7,     $5,     $2
+        addq    $1,     $24,    $24
+        addq    $28,    $23,    $28
+        cmpult  $28,    $23,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $213
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$213:
+        sll     $28,    32,     $23
+        ldq     $1,     8($18)
+        addq    $2,     $23,    $2
+        bis     $2,     $2,     $7
+        ldq     $4,     24($17)
+        addq    $22,    $7,     $22
+        srl     $1,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $25
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $0
+        srl     $28,    32,     $1
+        cmpult  $2,     $23,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $23
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $5
+        bis     $31,    1,      $21
+        addq    $25,    $5,     $25
+        cmpult  $25,    $5,     $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $217
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$217:
+        sll     $25,    32,     $5
+        ldq     $2,     16($18)
+        addq    $0,     $5,     $0
+        bis     $0,     $0,     $7
+        ldq     $4,     16($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $25,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $0,     $5,     $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $23,    $23
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $221
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$221:
+        sll     $28,    32,     $25
+        ldq     $2,     24($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     8($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $23,    $23
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $225
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$225:
+        sll     $0,     32,     $25
+        ldq     $2,     32($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     0($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $1,     $23,    $23
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $229
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$229:
+        sll     $28,    32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $22,    $7,     $22
+        stq     $22,    32($16)
+        ldq     $4,     0($17)
+        ldq     $5,     40($18)
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $22
+        mulq    $7,     $5,     $2
+        addq    $1,     $23,    $23
+        addq    $0,     $22,    $0
+        cmpult  $0,     $22,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $233
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$233:
+        sll     $0,     32,     $22
+        ldq     $1,     32($18)
+        addq    $2,     $22,    $2
+        bis     $2,     $2,     $7
+        ldq     $4,     8($17)
+        addq    $24,    $7,     $24
+        srl     $1,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $25
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $21
+        srl     $0,     32,     $1
+        cmpult  $2,     $22,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $22
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $5
+        bis     $31,    1,      $20
+        addq    $25,    $5,     $25
+        cmpult  $25,    $5,     $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $237
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$237:
+        sll     $25,    32,     $5
+        ldq     $2,     24($18)
+        addq    $21,    $5,     $21
+        bis     $21,    $21,    $7
+        ldq     $4,     16($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $25,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $21,    $5,     $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $22,    $22
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $241
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$241:
+        sll     $28,    32,     $25
+        ldq     $2,     16($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     24($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $22,    $22
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $245
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$245:
+        sll     $0,     32,     $25
+        ldq     $2,     8($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     32($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $22,    $22
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $249
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$249:
+        sll     $28,    32,     $25
+        ldq     $2,     0($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     40($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $1,     $22,    $22
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $253
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$253:
+        sll     $0,     32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $24,    $7,     $24
+        stq     $24,    40($16)
+        ldq     $4,     48($17)
+        ldq     $5,     0($18)
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $24
+        mulq    $7,     $5,     $2
+        addq    $1,     $22,    $22
+        addq    $28,    $24,    $28
+        cmpult  $28,    $24,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $257
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$257:
+        sll     $28,    32,     $24
+        ldq     $1,     8($18)
+        addq    $2,     $24,    $2
+        bis     $2,     $2,     $7
+        ldq     $4,     40($17)
+        addq    $23,    $7,     $23
+        srl     $1,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $25
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $0
+        srl     $28,    32,     $1
+        cmpult  $2,     $24,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $24
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $5
+        bis     $31,    1,      $21
+        addq    $25,    $5,     $25
+        cmpult  $25,    $5,     $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $261
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$261:
+        sll     $25,    32,     $5
+        ldq     $2,     16($18)
+        addq    $0,     $5,     $0
+        bis     $0,     $0,     $7
+        ldq     $4,     32($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $25,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $0,     $5,     $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $24,    $24
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $265
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$265:
+        sll     $28,    32,     $25
+        ldq     $2,     24($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     24($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $24,    $24
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $269
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$269:
+        sll     $0,     32,     $25
+        ldq     $2,     32($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     16($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $24,    $24
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $273
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$273:
+        sll     $28,    32,     $25
+        ldq     $2,     40($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     8($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $24,    $24
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $277
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$277:
+        sll     $0,     32,     $25
+        ldq     $2,     48($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     0($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $1,     $24,    $24
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $281
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$281:
+        sll     $28,    32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $23,    $7,     $23
+        stq     $23,    48($16)
+        ldq     $4,     0($17)
+        ldq     $5,     56($18)
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $23
+        mulq    $7,     $5,     $2
+        addq    $1,     $24,    $24
+        addq    $0,     $23,    $0
+        cmpult  $0,     $23,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $285
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$285:
+        sll     $0,     32,     $23
+        ldq     $1,     48($18)
+        addq    $2,     $23,    $2
+        bis     $2,     $2,     $7
+        ldq     $4,     8($17)
+        addq    $22,    $7,     $22
+        srl     $1,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $25
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $21
+        srl     $0,     32,     $1
+        cmpult  $2,     $23,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $23
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $5
+        bis     $31,    1,      $20
+        addq    $25,    $5,     $25
+        cmpult  $25,    $5,     $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $289
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$289:
+        sll     $25,    32,     $5
+        ldq     $2,     40($18)
+        addq    $21,    $5,     $21
+        bis     $21,    $21,    $7
+        ldq     $4,     16($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $25,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $21,    $5,     $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $23,    $23
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $293
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$293:
+        sll     $28,    32,     $25
+        ldq     $2,     32($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     24($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $23,    $23
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $297
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$297:
+        sll     $0,     32,     $25
+        ldq     $2,     24($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     32($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $23,    $23
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $301
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$301:
+        sll     $28,    32,     $25
+        ldq     $2,     16($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     40($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $23,    $23
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $305
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$305:
+        sll     $0,     32,     $25
+        ldq     $2,     8($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     48($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $23,    $23
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $309
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$309:
+        sll     $28,    32,     $25
+        ldq     $2,     0($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     56($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $1,     $23,    $23
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $313
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$313:
+        sll     $0,     32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $22,    $7,     $22
+        stq     $22,    56($16)
+        ldq     $4,     56($17)
+        ldq     $5,     8($18)
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $22
+        mulq    $7,     $5,     $2
+        addq    $1,     $23,    $23
+        addq    $28,    $22,    $28
+        cmpult  $28,    $22,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $317
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$317:
+        sll     $28,    32,     $22
+        ldq     $1,     16($18)
+        addq    $2,     $22,    $2
+        bis     $2,     $2,     $7
+        ldq     $4,     48($17)
+        addq    $24,    $7,     $24
+        srl     $1,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $25
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $0
+        srl     $28,    32,     $1
+        cmpult  $2,     $22,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $22
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $5
+        bis     $31,    1,      $21
+        addq    $25,    $5,     $25
+        cmpult  $25,    $5,     $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $321
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$321:
+        sll     $25,    32,     $5
+        ldq     $2,     24($18)
+        addq    $0,     $5,     $0
+        bis     $0,     $0,     $7
+        ldq     $4,     40($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $25,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $0,     $5,     $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $22,    $22
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $325
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$325:
+        sll     $28,    32,     $25
+        ldq     $2,     32($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     32($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $22,    $22
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $329
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$329:
+        sll     $0,     32,     $25
+        ldq     $2,     40($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     24($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $22,    $22
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $333
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$333:
+        sll     $28,    32,     $25
+        ldq     $2,     48($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     16($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $22,    $22
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $337
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$337:
+        sll     $0,     32,     $25
+        ldq     $2,     56($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     8($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $1,     $22,    $22
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $341
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$341:
+        sll     $28,    32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $24,    $7,     $24
+        stq     $24,    64($16)
+        ldq     $4,     16($17)
+        ldq     $5,     56($18)
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $24
+        mulq    $7,     $5,     $2
+        addq    $1,     $22,    $22
+        addq    $0,     $24,    $0
+        cmpult  $0,     $24,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $345
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$345:
+        sll     $0,     32,     $24
+        ldq     $1,     48($18)
+        addq    $2,     $24,    $2
+        bis     $2,     $2,     $7
+        ldq     $4,     24($17)
+        addq    $23,    $7,     $23
+        srl     $1,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $25
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $21
+        srl     $0,     32,     $1
+        cmpult  $2,     $24,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $24
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $5
+        bis     $31,    1,      $20
+        addq    $25,    $5,     $25
+        cmpult  $25,    $5,     $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $349
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$349:
+        sll     $25,    32,     $5
+        ldq     $2,     40($18)
+        addq    $21,    $5,     $21
+        bis     $21,    $21,    $7
+        ldq     $4,     32($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $25,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $21,    $5,     $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $24,    $24
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $353
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$353:
+        sll     $28,    32,     $25
+        ldq     $2,     32($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     40($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $24,    $24
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $357
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$357:
+        sll     $0,     32,     $25
+        ldq     $2,     24($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     48($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $24,    $24
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $361
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$361:
+        sll     $28,    32,     $25
+        ldq     $2,     16($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     56($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $1,     $24,    $24
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $365
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$365:
+        sll     $0,     32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $23,    $7,     $23
+        stq     $23,    72($16)
+        ldq     $4,     56($17)
+        ldq     $5,     24($18)
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $23
+        mulq    $7,     $5,     $2
+        addq    $1,     $24,    $24
+        addq    $28,    $23,    $28
+        cmpult  $28,    $23,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $369
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$369:
+        sll     $28,    32,     $23
+        ldq     $1,     32($18)
+        addq    $2,     $23,    $2
+        bis     $2,     $2,     $7
+        ldq     $4,     48($17)
+        addq    $22,    $7,     $22
+        srl     $1,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $25
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $0
+        srl     $28,    32,     $1
+        cmpult  $2,     $23,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $23
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $5
+        bis     $31,    1,      $21
+        addq    $25,    $5,     $25
+        cmpult  $25,    $5,     $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $373
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$373:
+        sll     $25,    32,     $5
+        ldq     $2,     40($18)
+        addq    $0,     $5,     $0
+        bis     $0,     $0,     $7
+        ldq     $4,     40($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $25,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $0,     $5,     $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $23,    $23
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $377
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$377:
+        sll     $28,    32,     $25
+        ldq     $2,     48($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     32($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $23,    $23
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $381
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$381:
+        sll     $0,     32,     $25
+        ldq     $2,     56($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     24($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $1,     $23,    $23
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $385
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$385:
+        sll     $28,    32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $22,    $7,     $22
+        stq     $22,    80($16)
+        ldq     $4,     32($17)
+        ldq     $5,     56($18)
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $22
+        mulq    $7,     $5,     $2
+        addq    $1,     $23,    $23
+        addq    $0,     $22,    $0
+        cmpult  $0,     $22,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $389
+        sll     $21,    32,     $1
+        addq    $6,     $1,     $6
+$389:
+        sll     $0,     32,     $22
+        ldq     $1,     48($18)
+        addq    $2,     $22,    $2
+        bis     $2,     $2,     $7
+        ldq     $4,     40($17)
+        addq    $24,    $7,     $24
+        srl     $1,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $25
+        zapnot  $1,     15,     $5
+        mulq    $7,     $5,     $21
+        srl     $0,     32,     $1
+        cmpult  $2,     $22,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $22
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $5
+        bis     $31,    1,      $20
+        addq    $25,    $5,     $25
+        cmpult  $25,    $5,     $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $393
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$393:
+        sll     $25,    32,     $5
+        ldq     $2,     40($18)
+        addq    $21,    $5,     $21
+        bis     $21,    $21,    $7
+        ldq     $4,     48($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $25,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $21,    $5,     $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $1,     $22,    $22
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $397
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$397:
+        sll     $28,    32,     $25
+        ldq     $2,     32($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     56($17)
+        addq    $24,    $7,     $24
+        srl     $2,     32,     $8
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $21
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $1,     $22,    $22
+        addq    $21,    $25,    $21
+        cmpult  $21,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $401
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$401:
+        sll     $21,    32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $24,    $7,     $24
+        stq     $24,    88($16)
+        ldq     $4,     56($17)
+        ldq     $5,     40($18)
+        cmpult  $24,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $0
+        srl     $21,    32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $23,    $6,     $23
+        cmpult  $23,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $24
+        mulq    $7,     $5,     $5
+        addq    $1,     $22,    $22
+        addq    $0,     $24,    $0
+        cmpult  $0,     $24,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $405
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$405:
+        sll     $0,     32,     $24
+        ldq     $2,     48($18)
+        addq    $5,     $24,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     48($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $24,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $24
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $5
+        addq    $28,    $25,    $28
+        cmpult  $28,    $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $409
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$409:
+        sll     $28,    32,     $25
+        ldq     $2,     56($18)
+        addq    $5,     $25,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     40($17)
+        addq    $23,    $7,     $23
+        srl     $2,     32,     $8
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $25,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $1,     $24,    $24
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $413
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$413:
+        sll     $0,     32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $23,    $7,     $23
+        stq     $23,    96($16)
+        ldq     $4,     48($17)
+        ldq     $5,     56($18)
+        cmpult  $23,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $22,    $6,     $22
+        cmpult  $22,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $23
+        mulq    $7,     $5,     $5
+        addq    $1,     $24,    $24
+        addq    $28,    $23,    $28
+        cmpult  $28,    $23,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $417
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$417:
+        sll     $28,    32,     $23
+        ldq     $2,     48($18)
+        addq    $5,     $23,    $5
+        bis     $5,     $5,     $7
+        ldq     $4,     56($17)
+        addq    $22,    $7,     $22
+        srl     $2,     32,     $8
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        mulq    $8,     $7,     $0
+        srl     $28,    32,     $1
+        addq    $6,     $1,     $6
+        cmpult  $5,     $23,    $1
+        zapnot  $2,     15,     $5
+        addq    $1,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $23
+        srl     $4,     32,     $6
+        mulq    $5,     $6,     $25
+        mulq    $7,     $5,     $2
+        addq    $0,     $25,    $0
+        cmpult  $0,     $25,    $1
+        mulq    $6,     $8,     $6
+        beq     $1,     $421
+        sll     $20,    32,     $1
+        addq    $6,     $1,     $6
+$421:
+        sll     $0,     32,     $25
+        addq    $2,     $25,    $2
+        bis     $2,     $2,     $7
+        addq    $22,    $7,     $22
+        stq     $22,    104($16)
+        ldq     $4,     56($17)
+        ldq     $5,     56($18)
+        cmpult  $22,    $7,     $3
+        zapnot  $4,     15,     $7
+        srl     $5,     32,     $8
+        mulq    $8,     $7,     $28
+        srl     $0,     32,     $1
+        cmpult  $2,     $25,    $2
+        addq    $6,     $1,     $6
+        addq    $2,     $6,     $6
+        addq    $3,     $6,     $6
+        addq    $24,    $6,     $24
+        cmpult  $24,    $6,     $1
+        srl     $4,     32,     $6
+        zapnot  $5,     15,     $5
+        mulq    $5,     $6,     $22
+        mulq    $7,     $5,     $2
+        addq    $1,     $23,    $23
+        addq    $28,    $22,    $28
+        cmpult  $28,    $22,    $1
+        mulq    $6,     $8,     $3
+        beq     $1,     $425
+        sll     $20,    32,     $1
+        addq    $3,     $1,     $3
+$425:
+        sll     $28,    32,     $22
+        srl     $28,    32,     $1
+        addq    $2,     $22,    $2
+        addq    $3,     $1,     $3
+        bis     $2,     $2,     $7
+        addq    $24,    $7,     $24
+        cmpult  $7,     $22,    $1
+        cmpult  $24,    $7,     $2
+        addq    $1,     $3,     $6
+        addq    $2,     $6,     $6
+        stq     $24,    112($16)
+        addq    $23,    $6,     $23
+        stq     $23,    120($16)
+        ret     $31,    ($26),  1
+        .end bn_mul_comba8
+        .text
+        .align 3
+        .globl bn_sqr_comba4
+        .ent bn_sqr_comba4
+bn_sqr_comba4:
+bn_sqr_comba4..ng:
+        .frame $30,0,$26,0
+        .prologue 0
+
+        ldq     $0,     0($17)
+        ldq     $1,     8($17)
+        ldq     $2,     16($17)
+        ldq     $3,     24($17)
+        bis     $31,    $31,    $6
+        mulq    $0,     $0,     $4
+        umulh   $0,     $0,     $5
+        stq     $4,     0($16)
+        bis     $31,    $31,    $4
+        mulq    $0,     $1,     $7
+        umulh   $0,     $1,     $8
+        cmplt   $7,     $31,    $22
+        cmplt   $8,     $31,    $23
+        addq    $7,     $7,     $7
+        addq    $8,     $8,     $8
+        addq    $8,     $22,    $8
+        addq    $4,     $23,    $4
+        addq    $5,     $7,     $5
+        addq    $6,     $8,     $6
+        cmpult  $5,     $7,     $24
+        cmpult  $6,     $8,     $25
+        addq    $6,     $24,    $6
+        addq    $4,     $25,    $4
+        stq     $5,     8($16)
+        bis     $31,    $31,    $5
+        mulq    $1,     $1,     $27
+        umulh   $1,     $1,     $28
+        addq    $6,     $27,    $6
+        addq    $4,     $28,    $4
+        cmpult  $6,     $27,    $21
+        cmpult  $4,     $28,    $20
+        addq    $4,     $21,    $4
+        addq    $5,     $20,    $5
+        mulq    $2,     $0,     $19
+        umulh   $2,     $0,     $18
+        cmplt   $19,    $31,    $17
+        cmplt   $18,    $31,    $22
+        addq    $19,    $19,    $19
+        addq    $18,    $18,    $18
+        addq    $18,    $17,    $18
+        addq    $5,     $22,    $5
+        addq    $6,     $19,    $6
+        addq    $4,     $18,    $4
+        cmpult  $6,     $19,    $23
+        cmpult  $4,     $18,    $7
+        addq    $4,     $23,    $4
+        addq    $5,     $7,     $5
+        stq     $6,     16($16)
+        bis     $31,    $31,    $6
+        mulq    $3,     $0,     $8
+        umulh   $3,     $0,     $24
+        cmplt   $8,     $31,    $25
+        cmplt   $24,    $31,    $27
+        addq    $8,     $8,     $8
+        addq    $24,    $24,    $24
+        addq    $24,    $25,    $24
+        addq    $6,     $27,    $6
+        addq    $4,     $8,     $4
+        addq    $5,     $24,    $5
+        cmpult  $4,     $8,     $28
+        cmpult  $5,     $24,    $21
+        addq    $5,     $28,    $5
+        addq    $6,     $21,    $6
+        mulq    $2,     $1,     $20
+        umulh   $2,     $1,     $17
+        cmplt   $20,    $31,    $22
+        cmplt   $17,    $31,    $19
+        addq    $20,    $20,    $20
+        addq    $17,    $17,    $17
+        addq    $17,    $22,    $17
+        addq    $6,     $19,    $6
+        addq    $4,     $20,    $4
+        addq    $5,     $17,    $5
+        cmpult  $4,     $20,    $18
+        cmpult  $5,     $17,    $23
+        addq    $5,     $18,    $5
+        addq    $6,     $23,    $6
+        stq     $4,     24($16)
+        bis     $31,    $31,    $4
+        mulq    $2,     $2,     $7
+        umulh   $2,     $2,     $25
+        addq    $5,     $7,     $5
+        addq    $6,     $25,    $6
+        cmpult  $5,     $7,     $27
+        cmpult  $6,     $25,    $8
+        addq    $6,     $27,    $6
+        addq    $4,     $8,     $4
+        mulq    $3,     $1,     $24
+        umulh   $3,     $1,     $28
+        cmplt   $24,    $31,    $21
+        cmplt   $28,    $31,    $22
+        addq    $24,    $24,    $24
+        addq    $28,    $28,    $28
+        addq    $28,    $21,    $28
+        addq    $4,     $22,    $4
+        addq    $5,     $24,    $5
+        addq    $6,     $28,    $6
+        cmpult  $5,     $24,    $19
+        cmpult  $6,     $28,    $20
+        addq    $6,     $19,    $6
+        addq    $4,     $20,    $4
+        stq     $5,     32($16)
+        bis     $31,    $31,    $5
+        mulq    $3,     $2,     $17
+        umulh   $3,     $2,     $18
+        cmplt   $17,    $31,    $23
+        cmplt   $18,    $31,    $7
+        addq    $17,    $17,    $17
+        addq    $18,    $18,    $18
+        addq    $18,    $23,    $18
+        addq    $5,     $7,     $5
+        addq    $6,     $17,    $6
+        addq    $4,     $18,    $4
+        cmpult  $6,     $17,    $25
+        cmpult  $4,     $18,    $27
+        addq    $4,     $25,    $4
+        addq    $5,     $27,    $5
+        stq     $6,     40($16)
+        bis     $31,    $31,    $6
+        mulq    $3,     $3,     $8
+        umulh   $3,     $3,     $21
+        addq    $4,     $8,     $4
+        addq    $5,     $21,    $5
+        cmpult  $4,     $8,     $22
+        cmpult  $5,     $21,    $24
+        addq    $5,     $22,    $5
+        addq    $6,     $24,    $6
+        stq     $4,     48($16)
+        stq     $5,     56($16)
+        ret     $31,($26),1
+        .end bn_sqr_comba4
+        .text
+        .align 3
+        .globl bn_sqr_comba8
+        .ent bn_sqr_comba8
+bn_sqr_comba8:
+bn_sqr_comba8..ng:
+        .frame $30,0,$26,0
+        .prologue 0
 
+        ldq     $0,     0($17)
+        ldq     $1,     8($17)
+        ldq     $2,     16($17)
+        ldq     $3,     24($17)
+        ldq     $4,     32($17)
+        ldq     $5,     40($17)
+        ldq     $6,     48($17)
+        ldq     $7,     56($17)
+        bis     $31,    $31,    $23
+        mulq    $0,     $0,     $8
+        umulh   $0,     $0,     $22
+        stq     $8,     0($16)
+        bis     $31,    $31,    $8
+        mulq    $1,     $0,     $24
+        umulh   $1,     $0,     $25
+        cmplt   $24,    $31,    $27
+        cmplt   $25,    $31,    $28
+        addq    $24,    $24,    $24
+        addq    $25,    $25,    $25
+        addq    $25,    $27,    $25
+        addq    $8,     $28,    $8
+        addq    $22,    $24,    $22
+        addq    $23,    $25,    $23
+        cmpult  $22,    $24,    $21
+        cmpult  $23,    $25,    $20
+        addq    $23,    $21,    $23
+        addq    $8,     $20,    $8
+        stq     $22,    8($16)
+        bis     $31,    $31,    $22
+        mulq    $1,     $1,     $19
+        umulh   $1,     $1,     $18
+        addq    $23,    $19,    $23
+        addq    $8,     $18,    $8
+        cmpult  $23,    $19,    $17
+        cmpult  $8,     $18,    $27
+        addq    $8,     $17,    $8
+        addq    $22,    $27,    $22
+        mulq    $2,     $0,     $28
+        umulh   $2,     $0,     $24
+        cmplt   $28,    $31,    $25
+        cmplt   $24,    $31,    $21
+        addq    $28,    $28,    $28
+        addq    $24,    $24,    $24
+        addq    $24,    $25,    $24
+        addq    $22,    $21,    $22
+        addq    $23,    $28,    $23
+        addq    $8,     $24,    $8
+        cmpult  $23,    $28,    $20
+        cmpult  $8,     $24,    $19
+        addq    $8,     $20,    $8
+        addq    $22,    $19,    $22
+        stq     $23,    16($16)
+        bis     $31,    $31,    $23
+        mulq    $2,     $1,     $18
+        umulh   $2,     $1,     $17
+        cmplt   $18,    $31,    $27
+        cmplt   $17,    $31,    $25
+        addq    $18,    $18,    $18
+        addq    $17,    $17,    $17
+        addq    $17,    $27,    $17
+        addq    $23,    $25,    $23
+        addq    $8,     $18,    $8
+        addq    $22,    $17,    $22
+        cmpult  $8,     $18,    $21
+        cmpult  $22,    $17,    $28
+        addq    $22,    $21,    $22
+        addq    $23,    $28,    $23
+        mulq    $3,     $0,     $24
+        umulh   $3,     $0,     $20
+        cmplt   $24,    $31,    $19
+        cmplt   $20,    $31,    $27
+        addq    $24,    $24,    $24
+        addq    $20,    $20,    $20
+        addq    $20,    $19,    $20
+        addq    $23,    $27,    $23
+        addq    $8,     $24,    $8
+        addq    $22,    $20,    $22
+        cmpult  $8,     $24,    $25
+        cmpult  $22,    $20,    $18
+        addq    $22,    $25,    $22
+        addq    $23,    $18,    $23
+        stq     $8,     24($16)
+        bis     $31,    $31,    $8
+        mulq    $2,     $2,     $17
+        umulh   $2,     $2,     $21
+        addq    $22,    $17,    $22
+        addq    $23,    $21,    $23
+        cmpult  $22,    $17,    $28
+        cmpult  $23,    $21,    $19
+        addq    $23,    $28,    $23
+        addq    $8,     $19,    $8
+        mulq    $3,     $1,     $27
+        umulh   $3,     $1,     $24
+        cmplt   $27,    $31,    $20
+        cmplt   $24,    $31,    $25
+        addq    $27,    $27,    $27
+        addq    $24,    $24,    $24
+        addq    $24,    $20,    $24
+        addq    $8,     $25,    $8
+        addq    $22,    $27,    $22
+        addq    $23,    $24,    $23
+        cmpult  $22,    $27,    $18
+        cmpult  $23,    $24,    $17
+        addq    $23,    $18,    $23
+        addq    $8,     $17,    $8
+        mulq    $4,     $0,     $21
+        umulh   $4,     $0,     $28
+        cmplt   $21,    $31,    $19
+        cmplt   $28,    $31,    $20
+        addq    $21,    $21,    $21
+        addq    $28,    $28,    $28
+        addq    $28,    $19,    $28
+        addq    $8,     $20,    $8
+        addq    $22,    $21,    $22
+        addq    $23,    $28,    $23
+        cmpult  $22,    $21,    $25
+        cmpult  $23,    $28,    $27
+        addq    $23,    $25,    $23
+        addq    $8,     $27,    $8
+        stq     $22,    32($16)
+        bis     $31,    $31,    $22
+        mulq    $3,     $2,     $24
+        umulh   $3,     $2,     $18
+        cmplt   $24,    $31,    $17
+        cmplt   $18,    $31,    $19
+        addq    $24,    $24,    $24
+        addq    $18,    $18,    $18
+        addq    $18,    $17,    $18
+        addq    $22,    $19,    $22
+        addq    $23,    $24,    $23
+        addq    $8,     $18,    $8
+        cmpult  $23,    $24,    $20
+        cmpult  $8,     $18,    $21
+        addq    $8,     $20,    $8
+        addq    $22,    $21,    $22
+        mulq    $4,     $1,     $28
+        umulh   $4,     $1,     $25
+        cmplt   $28,    $31,    $27
+        cmplt   $25,    $31,    $17
+        addq    $28,    $28,    $28
+        addq    $25,    $25,    $25
+        addq    $25,    $27,    $25
+        addq    $22,    $17,    $22
+        addq    $23,    $28,    $23
+        addq    $8,     $25,    $8
+        cmpult  $23,    $28,    $19
+        cmpult  $8,     $25,    $24
+        addq    $8,     $19,    $8
+        addq    $22,    $24,    $22
+        mulq    $5,     $0,     $18
+        umulh   $5,     $0,     $20
+        cmplt   $18,    $31,    $21
+        cmplt   $20,    $31,    $27
+        addq    $18,    $18,    $18
+        addq    $20,    $20,    $20
+        addq    $20,    $21,    $20
+        addq    $22,    $27,    $22
+        addq    $23,    $18,    $23
+        addq    $8,     $20,    $8
+        cmpult  $23,    $18,    $17
+        cmpult  $8,     $20,    $28
+        addq    $8,     $17,    $8
+        addq    $22,    $28,    $22
+        stq     $23,    40($16)
+        bis     $31,    $31,    $23
+        mulq    $3,     $3,     $25
+        umulh   $3,     $3,     $19
+        addq    $8,     $25,    $8
+        addq    $22,    $19,    $22
+        cmpult  $8,     $25,    $24
+        cmpult  $22,    $19,    $21
+        addq    $22,    $24,    $22
+        addq    $23,    $21,    $23
+        mulq    $4,     $2,     $27
+        umulh   $4,     $2,     $18
+        cmplt   $27,    $31,    $20
+        cmplt   $18,    $31,    $17
+        addq    $27,    $27,    $27
+        addq    $18,    $18,    $18
+        addq    $18,    $20,    $18
+        addq    $23,    $17,    $23
+        addq    $8,     $27,    $8
+        addq    $22,    $18,    $22
+        cmpult  $8,     $27,    $28
+        cmpult  $22,    $18,    $25
+        addq    $22,    $28,    $22
+        addq    $23,    $25,    $23
+        mulq    $5,     $1,     $19
+        umulh   $5,     $1,     $24
+        cmplt   $19,    $31,    $21
+        cmplt   $24,    $31,    $20
+        addq    $19,    $19,    $19
+        addq    $24,    $24,    $24
+        addq    $24,    $21,    $24
+        addq    $23,    $20,    $23
+        addq    $8,     $19,    $8
+        addq    $22,    $24,    $22
+        cmpult  $8,     $19,    $17
+        cmpult  $22,    $24,    $27
+        addq    $22,    $17,    $22
+        addq    $23,    $27,    $23
+        mulq    $6,     $0,     $18
+        umulh   $6,     $0,     $28
+        cmplt   $18,    $31,    $25
+        cmplt   $28,    $31,    $21
+        addq    $18,    $18,    $18
+        addq    $28,    $28,    $28
+        addq    $28,    $25,    $28
+        addq    $23,    $21,    $23
+        addq    $8,     $18,    $8
+        addq    $22,    $28,    $22
+        cmpult  $8,     $18,    $20
+        cmpult  $22,    $28,    $19
+        addq    $22,    $20,    $22
+        addq    $23,    $19,    $23
+        stq     $8,     48($16)
+        bis     $31,    $31,    $8
+        mulq    $4,     $3,     $24
+        umulh   $4,     $3,     $17
+        cmplt   $24,    $31,    $27
+        cmplt   $17,    $31,    $25
+        addq    $24,    $24,    $24
+        addq    $17,    $17,    $17
+        addq    $17,    $27,    $17
+        addq    $8,     $25,    $8
+        addq    $22,    $24,    $22
+        addq    $23,    $17,    $23
+        cmpult  $22,    $24,    $21
+        cmpult  $23,    $17,    $18
+        addq    $23,    $21,    $23
+        addq    $8,     $18,    $8
+        mulq    $5,     $2,     $28
+        umulh   $5,     $2,     $20
+        cmplt   $28,    $31,    $19
+        cmplt   $20,    $31,    $27
+        addq    $28,    $28,    $28
+        addq    $20,    $20,    $20
+        addq    $20,    $19,    $20
+        addq    $8,     $27,    $8
+        addq    $22,    $28,    $22
+        addq    $23,    $20,    $23
+        cmpult  $22,    $28,    $25
+        cmpult  $23,    $20,    $24
+        addq    $23,    $25,    $23
+        addq    $8,     $24,    $8
+        mulq    $6,     $1,     $17
+        umulh   $6,     $1,     $21
+        cmplt   $17,    $31,    $18
+        cmplt   $21,    $31,    $19
+        addq    $17,    $17,    $17
+        addq    $21,    $21,    $21
+        addq    $21,    $18,    $21
+        addq    $8,     $19,    $8
+        addq    $22,    $17,    $22
+        addq    $23,    $21,    $23
+        cmpult  $22,    $17,    $27
+        cmpult  $23,    $21,    $28
+        addq    $23,    $27,    $23
+        addq    $8,     $28,    $8
+        mulq    $7,     $0,     $20
+        umulh   $7,     $0,     $25
+        cmplt   $20,    $31,    $24
+        cmplt   $25,    $31,    $18
+        addq    $20,    $20,    $20
+        addq    $25,    $25,    $25
+        addq    $25,    $24,    $25
+        addq    $8,     $18,    $8
+        addq    $22,    $20,    $22
+        addq    $23,    $25,    $23
+        cmpult  $22,    $20,    $19
+        cmpult  $23,    $25,    $17
+        addq    $23,    $19,    $23
+        addq    $8,     $17,    $8
+        stq     $22,    56($16)
+        bis     $31,    $31,    $22
+        mulq    $4,     $4,     $21
+        umulh   $4,     $4,     $27
+        addq    $23,    $21,    $23
+        addq    $8,     $27,    $8
+        cmpult  $23,    $21,    $28
+        cmpult  $8,     $27,    $24
+        addq    $8,     $28,    $8
+        addq    $22,    $24,    $22
+        mulq    $5,     $3,     $18
+        umulh   $5,     $3,     $20
+        cmplt   $18,    $31,    $25
+        cmplt   $20,    $31,    $19
+        addq    $18,    $18,    $18
+        addq    $20,    $20,    $20
+        addq    $20,    $25,    $20
+        addq    $22,    $19,    $22
+        addq    $23,    $18,    $23
+        addq    $8,     $20,    $8
+        cmpult  $23,    $18,    $17
+        cmpult  $8,     $20,    $21
+        addq    $8,     $17,    $8
+        addq    $22,    $21,    $22
+        mulq    $6,     $2,     $27
+        umulh   $6,     $2,     $28
+        cmplt   $27,    $31,    $24
+        cmplt   $28,    $31,    $25
+        addq    $27,    $27,    $27
+        addq    $28,    $28,    $28
+        addq    $28,    $24,    $28
+        addq    $22,    $25,    $22
+        addq    $23,    $27,    $23
+        addq    $8,     $28,    $8
+        cmpult  $23,    $27,    $19
+        cmpult  $8,     $28,    $18
+        addq    $8,     $19,    $8
+        addq    $22,    $18,    $22
+        mulq    $7,     $1,     $20
+        umulh   $7,     $1,     $17
+        cmplt   $20,    $31,    $21
+        cmplt   $17,    $31,    $24
+        addq    $20,    $20,    $20
+        addq    $17,    $17,    $17
+        addq    $17,    $21,    $17
+        addq    $22,    $24,    $22
+        addq    $23,    $20,    $23
+        addq    $8,     $17,    $8
+        cmpult  $23,    $20,    $25
+        cmpult  $8,     $17,    $27
+        addq    $8,     $25,    $8
+        addq    $22,    $27,    $22
+        stq     $23,    64($16)
+        bis     $31,    $31,    $23
+        mulq    $5,     $4,     $28
+        umulh   $5,     $4,     $19
+        cmplt   $28,    $31,    $18
+        cmplt   $19,    $31,    $21
+        addq    $28,    $28,    $28
+        addq    $19,    $19,    $19
+        addq    $19,    $18,    $19
+        addq    $23,    $21,    $23
+        addq    $8,     $28,    $8
+        addq    $22,    $19,    $22
+        cmpult  $8,     $28,    $24
+        cmpult  $22,    $19,    $20
+        addq    $22,    $24,    $22
+        addq    $23,    $20,    $23
+        mulq    $6,     $3,     $17
+        umulh   $6,     $3,     $25
+        cmplt   $17,    $31,    $27
+        cmplt   $25,    $31,    $18
+        addq    $17,    $17,    $17
+        addq    $25,    $25,    $25
+        addq    $25,    $27,    $25
+        addq    $23,    $18,    $23
+        addq    $8,     $17,    $8
+        addq    $22,    $25,    $22
+        cmpult  $8,     $17,    $21
+        cmpult  $22,    $25,    $28
+        addq    $22,    $21,    $22
+        addq    $23,    $28,    $23
+        mulq    $7,     $2,     $19
+        umulh   $7,     $2,     $24
+        cmplt   $19,    $31,    $20
+        cmplt   $24,    $31,    $27
+        addq    $19,    $19,    $19
+        addq    $24,    $24,    $24
+        addq    $24,    $20,    $24
+        addq    $23,    $27,    $23
+        addq    $8,     $19,    $8
+        addq    $22,    $24,    $22
+        cmpult  $8,     $19,    $18
+        cmpult  $22,    $24,    $17
+        addq    $22,    $18,    $22
+        addq    $23,    $17,    $23
+        stq     $8,     72($16)
+        bis     $31,    $31,    $8
+        mulq    $5,     $5,     $25
+        umulh   $5,     $5,     $21
+        addq    $22,    $25,    $22
+        addq    $23,    $21,    $23
+        cmpult  $22,    $25,    $28
+        cmpult  $23,    $21,    $20
+        addq    $23,    $28,    $23
+        addq    $8,     $20,    $8
+        mulq    $6,     $4,     $27
+        umulh   $6,     $4,     $19
+        cmplt   $27,    $31,    $24
+        cmplt   $19,    $31,    $18
+        addq    $27,    $27,    $27
+        addq    $19,    $19,    $19
+        addq    $19,    $24,    $19
+        addq    $8,     $18,    $8
+        addq    $22,    $27,    $22
+        addq    $23,    $19,    $23
+        cmpult  $22,    $27,    $17
+        cmpult  $23,    $19,    $25
+        addq    $23,    $17,    $23
+        addq    $8,     $25,    $8
+        mulq    $7,     $3,     $21
+        umulh   $7,     $3,     $28
+        cmplt   $21,    $31,    $20
+        cmplt   $28,    $31,    $24
+        addq    $21,    $21,    $21
+        addq    $28,    $28,    $28
+        addq    $28,    $20,    $28
+        addq    $8,     $24,    $8
+        addq    $22,    $21,    $22
+        addq    $23,    $28,    $23
+        cmpult  $22,    $21,    $18
+        cmpult  $23,    $28,    $27
+        addq    $23,    $18,    $23
+        addq    $8,     $27,    $8
+        stq     $22,    80($16)
+        bis     $31,    $31,    $22
+        mulq    $6,     $5,     $19
+        umulh   $6,     $5,     $17
+        cmplt   $19,    $31,    $25
+        cmplt   $17,    $31,    $20
+        addq    $19,    $19,    $19
+        addq    $17,    $17,    $17
+        addq    $17,    $25,    $17
+        addq    $22,    $20,    $22
+        addq    $23,    $19,    $23
+        addq    $8,     $17,    $8
+        cmpult  $23,    $19,    $24
+        cmpult  $8,     $17,    $21
+        addq    $8,     $24,    $8
+        addq    $22,    $21,    $22
+        mulq    $7,     $4,     $28
+        umulh   $7,     $4,     $18
+        cmplt   $28,    $31,    $27
+        cmplt   $18,    $31,    $25
+        addq    $28,    $28,    $28
+        addq    $18,    $18,    $18
+        addq    $18,    $27,    $18
+        addq    $22,    $25,    $22
+        addq    $23,    $28,    $23
+        addq    $8,     $18,    $8
+        cmpult  $23,    $28,    $20
+        cmpult  $8,     $18,    $19
+        addq    $8,     $20,    $8
+        addq    $22,    $19,    $22
+        stq     $23,    88($16)
+        bis     $31,    $31,    $23
+        mulq    $6,     $6,     $17
+        umulh   $6,     $6,     $24
+        addq    $8,     $17,    $8
+        addq    $22,    $24,    $22
+        cmpult  $8,     $17,    $21
+        cmpult  $22,    $24,    $27
+        addq    $22,    $21,    $22
+        addq    $23,    $27,    $23
+        mulq    $7,     $5,     $25
+        umulh   $7,     $5,     $28
+        cmplt   $25,    $31,    $18
+        cmplt   $28,    $31,    $20
+        addq    $25,    $25,    $25
+        addq    $28,    $28,    $28
+        addq    $28,    $18,    $28
+        addq    $23,    $20,    $23
+        addq    $8,     $25,    $8
+        addq    $22,    $28,    $22
+        cmpult  $8,     $25,    $19
+        cmpult  $22,    $28,    $17
+        addq    $22,    $19,    $22
+        addq    $23,    $17,    $23
+        stq     $8,     96($16)
+        bis     $31,    $31,    $8
+        mulq    $7,     $6,     $24
+        umulh   $7,     $6,     $21
+        cmplt   $24,    $31,    $27
+        cmplt   $21,    $31,    $18
+        addq    $24,    $24,    $24
+        addq    $21,    $21,    $21
+        addq    $21,    $27,    $21
+        addq    $8,     $18,    $8
+        addq    $22,    $24,    $22
+        addq    $23,    $21,    $23
+        cmpult  $22,    $24,    $20
+        cmpult  $23,    $21,    $25
+        addq    $23,    $20,    $23
+        addq    $8,     $25,    $8
+        stq     $22,    104($16)
+        bis     $31,    $31,    $22
+        mulq    $7,     $7,     $28
+        umulh   $7,     $7,     $19
+        addq    $23,    $28,    $23
+        addq    $8,     $19,    $8
+        cmpult  $23,    $28,    $17
+        cmpult  $8,     $19,    $27
+        addq    $8,     $17,    $8
+        addq    $22,    $27,    $22
+        stq     $23,    112($16)
+        stq     $8,     120($16)
+        ret     $31,($26),1
+        .end bn_sqr_comba8
diff --git a/src/lib/libcrypto/bn/asm/mips3.s b/src/lib/libcrypto/bn/asm/mips3.s
index 191345d920..dca4105c7d 100644
--- a/src/lib/libcrypto/bn/asm/mips3.s
+++ b/src/lib/libcrypto/bn/asm/mips3.s
@@ -1,5 +1,5 @@
 .rdata
-.asciiz "mips3.s, Version 1.0"
+.asciiz "mips3.s, Version 1.1"
 .asciiz "MIPS III/IV ISA artwork by Andy Polyakov <appro@fy.chalmers.se>"
 
 /*
@@ -395,32 +395,32 @@ LEAF(bn_add_words)
 
 .L_bn_add_words_loop:
         ld      ta0,0(a2)
+        subu    a3,4
         ld      t1,8(a1)
-        ld      ta1,8(a2)
+        and     AT,a3,MINUS4
         ld      t2,16(a1)
-        ld      ta2,16(a2)
+        PTR_ADD a2,32
         ld      t3,24(a1)
-        ld      ta3,24(a2)
+        PTR_ADD a0,32
+        ld      ta1,-24(a2)
+        PTR_ADD a1,32
+        ld      ta2,-16(a2)
+        ld      ta3,-8(a2)
         daddu   ta0,t0
-        subu    a3,4
         sltu    t8,ta0,t0
         daddu   t0,ta0,v0
-        PTR_ADD a0,32
         sltu    v0,t0,ta0
         sd      t0,-32(a0)
         daddu   v0,t8
 
         daddu   ta1,t1
-        PTR_ADD a1,32
         sltu    t9,ta1,t1
         daddu   t1,ta1,v0
-        PTR_ADD a2,32
         sltu    v0,t1,ta1
         sd      t1,-24(a0)
         daddu   v0,t9
 
         daddu   ta2,t2
-        and     AT,a3,MINUS4
         sltu    t8,ta2,t2
         daddu   t2,ta2,v0
         sltu    v0,t2,ta2
@@ -495,25 +495,26 @@ LEAF(bn_sub_words)
 
 .L_bn_sub_words_loop:
         ld      ta0,0(a2)
+        subu    a3,4
         ld      t1,8(a1)
-        ld      ta1,8(a2)
+        and     AT,a3,MINUS4
         ld      t2,16(a1)
-        ld      ta2,16(a2)
+        PTR_ADD a2,32
         ld      t3,24(a1)
-        ld      ta3,24(a2)
+        PTR_ADD a0,32
+        ld      ta1,-24(a2)
+        PTR_ADD a1,32
+        ld      ta2,-16(a2)
+        ld      ta3,-8(a2)
         sltu    t8,t0,ta0
         dsubu   t0,ta0
-        subu    a3,4
         dsubu   ta0,t0,v0
-        and     AT,a3,MINUS4
-        sd      ta0,0(a0)
+        sd      ta0,-32(a0)
         MOVNZ   (t0,v0,t8)
 
         sltu    t9,t1,ta1
         dsubu   t1,ta1
-        PTR_ADD a0,32
         dsubu   ta1,t1,v0
-        PTR_ADD a1,32
         sd      ta1,-24(a0)
         MOVNZ   (t1,v0,t9)
 
@@ -521,7 +522,6 @@ LEAF(bn_sub_words)
         sltu    t8,t2,ta2
         dsubu   t2,ta2
         dsubu   ta2,t2,v0
-        PTR_ADD a2,32
         sd      ta2,-16(a0)
         MOVNZ   (t2,v0,t8)
 
@@ -574,6 +574,51 @@ END(bn_sub_words)
 
 #undef  MINUS4
 
+.align 5
+LEAF(bn_div_3_words)
+        .set    reorder
+        move    a3,a0           /* we know that bn_div_words doesn't
+                                 * touch a3, ta2, ta3 and preserves a2
+                                 * so that we can save two arguments
+                                 * and return address in registers
+                                 * instead of stack:-)
+                                 */
+        ld      a0,(a3)
+        move    ta2,a1
+        ld      a1,-8(a3)
+        bne     a0,a2,.L_bn_div_3_words_proceed
+        li      v0,-1
+        jr      ra
+.L_bn_div_3_words_proceed:
+        move    ta3,ra
+        bal     bn_div_words
+        move    ra,ta3
+        dmultu  ta2,v0
+        ld      t2,-16(a3)
+        move    ta0,zero
+        mfhi    t1
+        mflo    t0
+        sltu    t8,t1,v1
+.L_bn_div_3_words_inner_loop:
+        bnez    t8,.L_bn_div_3_words_inner_loop_done
+        sgeu    AT,t2,t0
+        seq     t9,t1,v1
+        and     AT,t9
+        sltu    t3,t0,ta2
+        daddu   v1,a2
+        dsubu   t1,t3
+        dsubu   t0,ta2
+        sltu    t8,t1,v1
+        sltu    ta0,v1,a2
+        or      t8,ta0
+        .set    noreorder
+        beqzl   AT,.L_bn_div_3_words_inner_loop
+        dsubu   v0,1
+        .set    reorder
+.L_bn_div_3_words_inner_loop_done:
+        jr      ra
+END(bn_div_3_words)
+
 .align  5
 LEAF(bn_div_words)
         .set    noreorder
@@ -633,16 +678,16 @@ LEAF(bn_div_words)
         seq     t8,HH,t1
         sltu    AT,HH,t1
         and     t2,t8
+        sltu    v0,t0,a2
         or      AT,t2
         .set    noreorder
         beqz    AT,.L_bn_div_words_inner_loop1_done
-        sltu    t2,t0,a2
-        .set    reorder
-        dsubu   QT,1
+        dsubu   t1,v0
         dsubu   t0,a2
-        dsubu   t1,t2
         b       .L_bn_div_words_inner_loop1
-.L_bn_div_words_inner_loop1_done:       
+        dsubu   QT,1
+        .set    reorder
+.L_bn_div_words_inner_loop1_done:
 
         dsll    a1,32
         dsubu   a0,t3,t0
@@ -655,6 +700,7 @@ LEAF(bn_div_words)
         ddivu   zero,a0,DH
         mflo    QT
 .L_bn_div_words_skip_div2:
+#undef  DH
         dmultu  a2,QT
         dsll    t3,a0,32
         dsrl    AT,a1,32
@@ -666,69 +712,26 @@ LEAF(bn_div_words)
         seq     t8,HH,t1
         sltu    AT,HH,t1
         and     t2,t8
+        sltu    v1,t0,a2
         or      AT,t2
         .set    noreorder
         beqz    AT,.L_bn_div_words_inner_loop2_done
-        sltu    t2,t0,a2
-        .set    reorder
-        dsubu   QT,1
+        dsubu   t1,v1
         dsubu   t0,a2
-        dsubu   t1,t2
         b       .L_bn_div_words_inner_loop2
+        dsubu   QT,1
+        .set    reorder
 .L_bn_div_words_inner_loop2_done:       
+#undef  HH
 
         dsubu   a0,t3,t0
         or      v0,QT
         dsrl    v1,a0,t9        /* v1 contains remainder if anybody wants it */
         dsrl    a2,t9           /* restore a2 */
         jr      ra
-#undef  HH
-#undef  DH
 #undef  QT
 END(bn_div_words)
 
-.align 5
-LEAF(bn_div_3_words)
-        .set    reorder
-        move    a3,a0           /* we know that bn_div_words doesn't
-                                 * touch a3, ta2, ta3 and preserves a2
-                                 * so that we can save two arguments
-                                 * and return address in registers
-                                 * instead of stack:-)
-                                 */
-        ld      a0,(a3)
-        move    ta2,a2
-        move    a2,a1
-        ld      a1,-8(a3)
-        move    ta3,ra
-        move    v1,zero
-        li      v0,-1
-        beq     a0,a2,.L_bn_div_3_words_skip_div
-        jal     bn_div_words
-        move    ra,ta3
-.L_bn_div_3_words_skip_div:
-        dmultu  ta2,v0
-        ld      t2,-16(a3)
-        mflo    t0
-        mfhi    t1
-.L_bn_div_3_words_inner_loop:
-        sgeu    AT,t2,t0
-        seq     t9,t1,v1
-        sltu    t8,t1,v1
-        and     AT,t9
-        or      AT,t8
-        bnez    AT,.L_bn_div_3_words_inner_loop_done
-        daddu   v1,a2
-        sltu    t3,t0,ta2
-        sltu    AT,v1,a2
-        dsubu   v0,1
-        dsubu   t0,ta2
-        dsubu   t1,t3
-        beqz    AT,.L_bn_div_3_words_inner_loop
-.L_bn_div_3_words_inner_loop_done:
-        jr      ra
-END(bn_div_3_words)
-
 #define a_0     t0
 #define a_1     t1
 #define a_2     t2
@@ -846,6 +849,7 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
+        sltu    c_3,c_2,t_2
         dmultu  a_1,b_2         /* mul_add_c(a[1],b[2],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
@@ -853,7 +857,8 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
-        sltu    c_3,c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
         dmultu  a_2,b_1         /* mul_add_c(a[2],b[1],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
@@ -881,6 +886,7 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
+        sltu    c_1,c_3,t_2
         dmultu  a_3,b_1         /* mul_add_c(a[3],b[1],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
@@ -888,7 +894,8 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
         dmultu  a_2,b_2         /* mul_add_c(a[2],b[2],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
@@ -925,6 +932,7 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
+        sltu    c_2,c_1,t_2
         dmultu  a_1,b_4         /* mul_add_c(a[1],b[4],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
@@ -932,7 +940,8 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
-        sltu    c_2,c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
         dmultu  a_2,b_3         /* mul_add_c(a[2],b[3],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
@@ -978,6 +987,7 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
+        sltu    c_3,c_2,t_2
         dmultu  a_5,b_1         /* mul_add_c(a[5],b[1],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
@@ -985,7 +995,8 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
-        sltu    c_3,c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
         dmultu  a_4,b_2         /* mul_add_c(a[4],b[2],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
@@ -1040,6 +1051,7 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
+        sltu    c_1,c_3,t_2
         dmultu  a_1,b_6         /* mul_add_c(a[1],b[6],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
@@ -1047,7 +1059,8 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
         dmultu  a_2,b_5         /* mul_add_c(a[2],b[5],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
@@ -1111,6 +1124,7 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
+        sltu    c_2,c_1,t_2
         dmultu  a_6,b_2         /* mul_add_c(a[6],b[2],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
@@ -1118,7 +1132,8 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
-        sltu    c_2,c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
         dmultu  a_5,b_3         /* mul_add_c(a[5],b[3],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
@@ -1173,6 +1188,7 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
+        sltu    c_3,c_2,t_2
         dmultu  a_3,b_6         /* mul_add_c(a[3],b[6],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
@@ -1180,7 +1196,8 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
-        sltu    c_3,c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
         dmultu  a_4,b_5         /* mul_add_c(a[4],b[5],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
@@ -1226,6 +1243,7 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
+        sltu    c_1,c_3,t_2
         dmultu  a_6,b_4         /* mul_add_c(a[6],b[4],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
@@ -1233,7 +1251,8 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
         dmultu  a_5,b_5         /* mul_add_c(a[5],b[5],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
@@ -1270,6 +1289,7 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
+        sltu    c_2,c_1,t_2
         dmultu  a_5,b_6         /* mul_add_c(a[5],b[6],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
@@ -1277,7 +1297,8 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
-        sltu    c_2,c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
         dmultu  a_6,b_5         /* mul_add_c(a[6],b[5],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
@@ -1305,6 +1326,7 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
+        sltu    c_3,c_2,t_2
         dmultu  a_6,b_6         /* mul_add_c(a[6],b[6],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
@@ -1312,7 +1334,8 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
-        sltu    c_3,c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
         dmultu  a_5,b_7         /* mul_add_c(a[5],b[7],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
@@ -1331,6 +1354,7 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
+        sltu    c_1,c_3,t_2
         dmultu  a_7,b_6         /* mul_add_c(a[7],b[6],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
@@ -1338,7 +1362,8 @@ LEAF(bn_mul_comba8)
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
         sd      c_2,104(a0)     /* r[13]=c2; */
 
         dmultu  a_7,b_7         /* mul_add_c(a[7],b[7],c3,c1,c2); */
@@ -1427,6 +1452,7 @@ LEAF(bn_mul_comba4)
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
+        sltu    c_3,c_2,t_2
         dmultu  a_1,b_2         /* mul_add_c(a[1],b[2],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
@@ -1434,7 +1460,8 @@ LEAF(bn_mul_comba4)
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
-        sltu    c_3,c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
         dmultu  a_2,b_1         /* mul_add_c(a[2],b[1],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
@@ -1462,6 +1489,7 @@ LEAF(bn_mul_comba4)
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
+        sltu    c_1,c_3,t_2
         dmultu  a_2,b_2         /* mul_add_c(a[2],b[2],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
@@ -1469,7 +1497,8 @@ LEAF(bn_mul_comba4)
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
         dmultu  a_1,b_3         /* mul_add_c(a[1],b[3],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
@@ -1488,6 +1517,7 @@ LEAF(bn_mul_comba4)
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
+        sltu    c_2,c_1,t_2
         dmultu  a_3,b_2         /* mul_add_c(a[3],b[2],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
@@ -1495,7 +1525,8 @@ LEAF(bn_mul_comba4)
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
-        sltu    c_2,c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
         sd      c_3,40(a0)
 
         dmultu  a_3,b_3         /* mul_add_c(a[3],b[3],c1,c2,c3); */
@@ -1540,28 +1571,30 @@ LEAF(bn_sqr_comba8)
         dmultu  a_0,a_1         /* mul_add_c2(a[0],b[1],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
+        slt     c_1,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   c_3,t_2,AT
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   t_2,AT
-        daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
         sd      c_2,8(a0)
 
         dmultu  a_2,a_0         /* mul_add_c2(a[2],b[0],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_3,t_1
-        sltu    AT,c_3,t_1
-        daddu   a2,t_2,AT
-        daddu   c_1,a2
+        slt     c_2,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_3,t_1
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
-        sltu    c_2,c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
         dmultu  a_1,a_1         /* mul_add_c(a[1],b[1],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
@@ -1576,24 +1609,26 @@ LEAF(bn_sqr_comba8)
         dmultu  a_0,a_3         /* mul_add_c2(a[0],b[3],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_1,t_1
-        sltu    AT,c_1,t_1
-        daddu   a2,t_2,AT
-        daddu   c_2,a2
+        slt     c_3,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_1,t_1
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
-        sltu    c_3,c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
         dmultu  a_1,a_2         /* mul_add_c2(a[1],b[2],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_1,t_1
-        sltu    AT,c_1,t_1
-        daddu   a2,t_2,AT
-        daddu   c_2,a2
-        sltu    AT,c_2,a2
+        slt     AT,t_2,zero
         daddu   c_3,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_1,t_1
         sltu    AT,c_1,t_1
         daddu   t_2,AT
@@ -1605,24 +1640,26 @@ LEAF(bn_sqr_comba8)
         dmultu  a_4,a_0         /* mul_add_c2(a[4],b[0],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   a2,t_2,AT
-        daddu   c_3,a2
+        slt     c_1,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
         dmultu  a_3,a_1         /* mul_add_c2(a[3],b[1],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   a2,t_2,AT
-        daddu   c_3,a2
-        sltu    AT,c_3,a2
+        slt     AT,t_2,zero
         daddu   c_1,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   t_2,AT
@@ -1643,24 +1680,26 @@ LEAF(bn_sqr_comba8)
         dmultu  a_0,a_5         /* mul_add_c2(a[0],b[5],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_3,t_1
-        sltu    AT,c_3,t_1
-        daddu   a2,t_2,AT
-        daddu   c_1,a2
+        slt     c_2,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_3,t_1
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
-        sltu    c_2,c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
         dmultu  a_1,a_4         /* mul_add_c2(a[1],b[4],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_3,t_1
-        sltu    AT,c_3,t_1
-        daddu   a2,t_2,AT
-        daddu   c_1,a2
-        sltu    AT,c_1,a2
+        slt     AT,t_2,zero
         daddu   c_2,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_3,t_1
         sltu    AT,c_3,t_1
         daddu   t_2,AT
@@ -1670,12 +1709,12 @@ LEAF(bn_sqr_comba8)
         dmultu  a_2,a_3         /* mul_add_c2(a[2],b[3],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_3,t_1
-        sltu    AT,c_3,t_1
-        daddu   a2,t_2,AT
-        daddu   c_1,a2
-        sltu    AT,c_1,a2
+        slt     AT,t_2,zero
         daddu   c_2,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_3,t_1
         sltu    AT,c_3,t_1
         daddu   t_2,AT
@@ -1687,24 +1726,26 @@ LEAF(bn_sqr_comba8)
         dmultu  a_6,a_0         /* mul_add_c2(a[6],b[0],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_1,t_1
-        sltu    AT,c_1,t_1
-        daddu   a2,t_2,AT
-        daddu   c_2,a2
+        slt     c_3,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_1,t_1
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
-        sltu    c_3,c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
         dmultu  a_5,a_1         /* mul_add_c2(a[5],b[1],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_1,t_1
-        sltu    AT,c_1,t_1
-        daddu   a2,t_2,AT
-        daddu   c_2,a2
-        sltu    AT,c_2,a2
+        slt     AT,t_2,zero
         daddu   c_3,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_1,t_1
         sltu    AT,c_1,t_1
         daddu   t_2,AT
@@ -1714,12 +1755,12 @@ LEAF(bn_sqr_comba8)
         dmultu  a_4,a_2         /* mul_add_c2(a[4],b[2],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_1,t_1
-        sltu    AT,c_1,t_1
-        daddu   a2,t_2,AT
-        daddu   c_2,a2
-        sltu    AT,c_2,a2
+        slt     AT,t_2,zero
         daddu   c_3,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_1,t_1
         sltu    AT,c_1,t_1
         daddu   t_2,AT
@@ -1740,24 +1781,26 @@ LEAF(bn_sqr_comba8)
         dmultu  a_0,a_7         /* mul_add_c2(a[0],b[7],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   a2,t_2,AT
-        daddu   c_3,a2
+        slt     c_1,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
         dmultu  a_1,a_6         /* mul_add_c2(a[1],b[6],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   a2,t_2,AT
-        daddu   c_3,a2
-        sltu    AT,c_3,a2
+        slt     AT,t_2,zero
         daddu   c_1,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   t_2,AT
@@ -1767,12 +1810,12 @@ LEAF(bn_sqr_comba8)
         dmultu  a_2,a_5         /* mul_add_c2(a[2],b[5],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   a2,t_2,AT
-        daddu   c_3,a2
-        sltu    AT,c_3,a2
+        slt     AT,t_2,zero
         daddu   c_1,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   t_2,AT
@@ -1782,12 +1825,12 @@ LEAF(bn_sqr_comba8)
         dmultu  a_3,a_4         /* mul_add_c2(a[3],b[4],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   a2,t_2,AT
-        daddu   c_3,a2
-        sltu    AT,c_3,a2
+        slt     AT,t_2,zero
         daddu   c_1,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   t_2,AT
@@ -1799,24 +1842,26 @@ LEAF(bn_sqr_comba8)
         dmultu  a_7,a_1         /* mul_add_c2(a[7],b[1],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_3,t_1
-        sltu    AT,c_3,t_1
-        daddu   a2,t_2,AT
-        daddu   c_1,a2
+        slt     c_2,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_3,t_1
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
-        sltu    c_2,c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
         dmultu  a_6,a_2         /* mul_add_c2(a[6],b[2],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_3,t_1
-        sltu    AT,c_3,t_1
-        daddu   a2,t_2,AT
-        daddu   c_1,a2
-        sltu    AT,c_1,a2
+        slt     AT,t_2,zero
         daddu   c_2,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_3,t_1
         sltu    AT,c_3,t_1
         daddu   t_2,AT
@@ -1826,12 +1871,12 @@ LEAF(bn_sqr_comba8)
         dmultu  a_5,a_3         /* mul_add_c2(a[5],b[3],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_3,t_1
-        sltu    AT,c_3,t_1
-        daddu   a2,t_2,AT
-        daddu   c_1,a2
-        sltu    AT,c_1,a2
+        slt     AT,t_2,zero
         daddu   c_2,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_3,t_1
         sltu    AT,c_3,t_1
         daddu   t_2,AT
@@ -1852,24 +1897,26 @@ LEAF(bn_sqr_comba8)
         dmultu  a_2,a_7         /* mul_add_c2(a[2],b[7],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_1,t_1
-        sltu    AT,c_1,t_1
-        daddu   a2,t_2,AT
-        daddu   c_2,a2
+        slt     c_3,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_1,t_1
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
-        sltu    c_3,c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
         dmultu  a_3,a_6         /* mul_add_c2(a[3],b[6],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_1,t_1
-        sltu    AT,c_1,t_1
-        daddu   a2,t_2,AT
-        daddu   c_2,a2
-        sltu    AT,c_2,a2
+        slt     AT,t_2,zero
         daddu   c_3,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_1,t_1
         sltu    AT,c_1,t_1
         daddu   t_2,AT
@@ -1879,12 +1926,12 @@ LEAF(bn_sqr_comba8)
         dmultu  a_4,a_5         /* mul_add_c2(a[4],b[5],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_1,t_1
-        sltu    AT,c_1,t_1
-        daddu   a2,t_2,AT
-        daddu   c_2,a2
-        sltu    AT,c_2,a2
+        slt     AT,t_2,zero
         daddu   c_3,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_1,t_1
         sltu    AT,c_1,t_1
         daddu   t_2,AT
@@ -1896,24 +1943,26 @@ LEAF(bn_sqr_comba8)
         dmultu  a_7,a_3         /* mul_add_c2(a[7],b[3],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   a2,t_2,AT
-        daddu   c_3,a2
+        slt     c_1,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
         dmultu  a_6,a_4         /* mul_add_c2(a[6],b[4],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   a2,t_2,AT
-        daddu   c_3,a2
-        sltu    AT,c_3,a2
+        slt     AT,t_2,zero
         daddu   c_1,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   t_2,AT
@@ -1934,24 +1983,26 @@ LEAF(bn_sqr_comba8)
         dmultu  a_4,a_7         /* mul_add_c2(a[4],b[7],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_3,t_1
-        sltu    AT,c_3,t_1
-        daddu   a2,t_2,AT
-        daddu   c_1,a2
+        slt     c_2,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_3,t_1
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
-        sltu    c_2,c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
         dmultu  a_5,a_6         /* mul_add_c2(a[5],b[6],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_3,t_1
-        sltu    AT,c_3,t_1
-        daddu   a2,t_2,AT
-        daddu   c_1,a2
-        sltu    AT,c_1,a2
+        slt     AT,t_2,zero
         daddu   c_2,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_3,t_1
         sltu    AT,c_3,t_1
         daddu   t_2,AT
@@ -1963,15 +2014,17 @@ LEAF(bn_sqr_comba8)
         dmultu  a_7,a_5         /* mul_add_c2(a[7],b[5],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_1,t_1
-        sltu    AT,c_1,t_1
-        daddu   a2,t_2,AT
-        daddu   c_2,a2
+        slt     c_3,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_1,t_1
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
-        sltu    c_3,c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
         dmultu  a_6,a_6         /* mul_add_c(a[6],b[6],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
@@ -1986,15 +2039,17 @@ LEAF(bn_sqr_comba8)
         dmultu  a_6,a_7         /* mul_add_c2(a[6],b[7],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   a2,t_2,AT
-        daddu   c_3,a2
+        slt     c_1,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
         sd      c_2,104(a0)
 
         dmultu  a_7,a_7         /* mul_add_c(a[7],b[7],c3,c1,c2); */
@@ -2025,28 +2080,30 @@ LEAF(bn_sqr_comba4)
         dmultu  a_0,a_1         /* mul_add_c2(a[0],b[1],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
+        slt     c_1,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   c_3,t_2,AT
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   t_2,AT
-        daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
         sd      c_2,8(a0)
 
         dmultu  a_2,a_0         /* mul_add_c2(a[2],b[0],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_3,t_1
-        sltu    AT,c_3,t_1
-        daddu   a2,t_2,AT
-        daddu   c_1,a2
+        slt     c_2,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_3,t_1
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
-        sltu    c_2,c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
         dmultu  a_1,a_1         /* mul_add_c(a[1],b[1],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
@@ -2061,24 +2118,26 @@ LEAF(bn_sqr_comba4)
         dmultu  a_0,a_3         /* mul_add_c2(a[0],b[3],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_1,t_1
-        sltu    AT,c_1,t_1
-        daddu   a2,t_2,AT
-        daddu   c_2,a2
+        slt     c_3,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_1,t_1
         sltu    AT,c_1,t_1
         daddu   t_2,AT
         daddu   c_2,t_2
-        sltu    c_3,c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
         dmultu  a_1,a_2         /* mul_add_c(a2[1],b[2],c1,c2,c3); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_1,t_1
-        sltu    AT,c_1,t_1
-        daddu   a2,t_2,AT
-        daddu   c_2,a2
-        sltu    AT,c_2,a2
+        slt     AT,t_2,zero
         daddu   c_3,AT
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_1,t_1
         sltu    AT,c_1,t_1
         daddu   t_2,AT
@@ -2090,15 +2149,17 @@ LEAF(bn_sqr_comba4)
         dmultu  a_3,a_1         /* mul_add_c2(a[3],b[1],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_2,t_1
-        sltu    AT,c_2,t_1
-        daddu   a2,t_2,AT
-        daddu   c_3,a2
+        slt     c_1,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_2,t_1
         sltu    AT,c_2,t_1
         daddu   t_2,AT
         daddu   c_3,t_2
-        sltu    c_1,c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
         dmultu  a_2,a_2         /* mul_add_c(a[2],b[2],c2,c3,c1); */
         mflo    t_1
         mfhi    t_2
@@ -2113,15 +2174,17 @@ LEAF(bn_sqr_comba4)
         dmultu  a_2,a_3         /* mul_add_c2(a[2],b[3],c3,c1,c2); */
         mflo    t_1
         mfhi    t_2
-        daddu   c_3,t_1
-        sltu    AT,c_3,t_1
-        daddu   a2,t_2,AT
-        daddu   c_1,a2
+        slt     c_2,t_2,zero
+        dsll    t_2,1
+        slt     a2,t_1,zero
+        daddu   t_2,a2
+        dsll    t_1,1
         daddu   c_3,t_1
         sltu    AT,c_3,t_1
         daddu   t_2,AT
         daddu   c_1,t_2
-        sltu    c_2,c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
         sd      c_3,40(a0)
 
         dmultu  a_3,a_3         /* mul_add_c(a[3],b[3],c1,c2,c3); */
diff --git a/src/lib/libcrypto/bn/asm/vms.mar b/src/lib/libcrypto/bn/asm/vms.mar
index ac9d57d7b0..465f2774b6 100644
--- a/src/lib/libcrypto/bn/asm/vms.mar
+++ b/src/lib/libcrypto/bn/asm/vms.mar
@@ -162,442 +162,237 @@ n=12 ;(AP)	n	by value (input)
         movl    #1,r0                   ; return SS$_NORMAL
         ret
 
-        .title  (generated)
-
-        .psect  code,nowrt
-
-.entry  BN_DIV_WORDS,^m<r2,r3,r4,r5,r6,r7,r8,r9,r10>
-        subl2   #4,sp
-
-        clrl    r9
-        movl    #2,r8
-
-        tstl    12(ap)
-        bneq    noname.2
-        mnegl   #1,r10
-        brw     noname.3
-        tstl    r0
-        nop     
-noname.2:
-
-        pushl   12(ap)
-        calls   #1,BN_NUM_BITS_WORD
-        movl    r0,r7
-
-        cmpl    r7,#32
-        beql    noname.4
-        ashl    r7,#1,r2
-        cmpl    4(ap),r2
-        blequ   noname.4
-
-        pushl   r7
-        calls   #1,BN_DIV_WORDS_ABORT
-noname.4:
-
-        subl3   r7,#32,r7
-
-        movl    12(ap),r2
-        cmpl    4(ap),r2
-        blssu   noname.5
-        subl2   r2,4(ap)
-noname.5:
-
-        tstl    r7
-        beql    noname.6
-
-        ashl    r7,r2,12(ap)
-
-        ashl    r7,4(ap),r4
-        subl3   r7,#32,r3
-        subl3   r3,#32,r2
-        extzv   r3,r2,8(ap),r2
-        bisl3   r4,r2,4(ap)
-
-        ashl    r7,8(ap),8(ap)
-noname.6:
-
-        bicl3   #65535,12(ap),r2
-        extzv   #16,#16,r2,r5
-
-        bicl3   #-65536,12(ap),r6
-
-noname.7:
-
-        moval   4(ap),r2
-        movzwl  2(r2),r0
-        cmpl    r0,r5
-        bneq    noname.8
-
-        movzwl  #65535,r4
-        brb     noname.9
-noname.8:
-
-        clrl    r1
-        movl    (r2),r0
-        movl    r5,r2
-        bgeq    vcg.1
-        cmpl    r2,r0
-        bgtru   vcg.2
-        incl    r1
-        brb     vcg.2
-        nop     
-vcg.1:
-        ediv    r2,r0,r1,r0
-vcg.2:
-        movl    r1,r4
-noname.9:
-
-noname.10:
-
-        mull3   r5,r4,r0
-        subl3   r0,4(ap),r3
-
-        bicl3   #65535,r3,r0
-        bneq    noname.13
-        mull3   r6,r4,r2
-        ashl    #16,r3,r1
-        bicl3   #65535,8(ap),r0
-        extzv   #16,#16,r0,r0
-        addl2   r0,r1
-        cmpl    r2,r1
-        bgtru   noname.12
-noname.11:
-
-        brb     noname.13
-        nop     
-noname.12:
-
-        decl    r4
-        brb     noname.10
-noname.13:
-
-        mull3   r5,r4,r1
-
-        mull3   r6,r4,r0
-
-        extzv   #16,#16,r0,r3
-
-        ashl    #16,r0,r2
-        bicl3   #65535,r2,r0
-
-        addl2   r3,r1
-
-        moval   8(ap),r3
-        cmpl    (r3),r0
-        bgequ   noname.15
-        incl    r1
-noname.15:
-
-        subl2   r0,(r3)
-
-        cmpl    4(ap),r1
-        bgequ   noname.16
-
-        addl2   12(ap),4(ap)
-
-        decl    r4
-noname.16:
-
-        subl2   r1,4(ap)
-
-        decl    r8
-        beql    noname.18
-noname.17:
-
-        ashl    #16,r4,r9
+        .title  vax_bn_div_words  unsigned divide
+;
+; Richard Levitte 20-Nov-2000
+;
+; ULONG bn_div_words(ULONG h, ULONG l, ULONG d)
+; {
+;       return ((ULONG)((((ULLONG)h)<<32)|l) / (ULLONG)d);
+; }
+;
+; Using EDIV would be very easy, if it didn't do signed calculations.
+; Therefore, som extra things have to happen around it.  The way to
+; handle that is to shift all operands right one step (basically dividing
+; them by 2) and handle the different cases depending on what the lowest
+; bit of each operand was.
+;
+; To start with, let's define the following:
+;
+; a' = l & 1
+; a2 = <h,l> >> 1       # UNSIGNED shift!
+; b' = d & 1
+; b2 = d >> 1           # UNSIGNED shift!
+;
+; Now, use EDIV to calculate a quotient and a remainder:
+;
+; q'' = a2/b2
+; r'' = a2 - q''*b2
+;
+; If b' is 0, the quotient is already correct, we just need to adjust the
+; remainder:
+;
+; if (b' == 0)
+;   {
+;     r = 2*r'' + a'
+;     q = q''
+;   }
+;
+; If b' is 1, we need to do other adjustements.  The first thought is the
+; following (note that r' will not always have the right value, but an
+; adjustement follows further down):
+;
+; if (b' == 1)
+;   {
+;     q' = q''
+;     r' = a - q'*b
+;
+; However, one can note the folowing relationship:
+;
+;                         r'' = a2 - q''*b2
+;                  =>   2*r'' = 2*a2 - 2*q''*b2
+;                             = { a = 2*a2 + a', b = 2*b2 + b' = 2*b2 + 1,
+;                                 q' = q'' }
+;                             = a - a' - q'*(b - 1)
+;                             = a - q'*b - a' + q'
+;                             = r' - a' + q'
+;                  =>     r'  = 2*r'' - q' + a'
+;
+; This enables us to use r'' instead of discarding and calculating another
+; modulo:
+;
+; if (b' == 1)
+;   {
+;     q' = q''
+;     r' = (r'' << 1) - q' + a'
+;
+; Now, all we have to do is adjust r', because it might be < 0:
+;
+;     while (r' < 0)
+;       {
+;         r' = r' + b
+;         q' = q' - 1
+;       }
+;   }
+;
+; return q'
 
-        ashl    #16,4(ap),r2
-        movzwl  2(r3),r0
-        bisl2   r0,r2
-        bicl3   #0,r2,4(ap)
+h=4 ;(AP)       h       by value (input)
+l=8 ;(AP)       l       by value (input)
+d=12 ;(AP)      d       by value (input)
 
-        bicl3   #-65536,(r3),r0
-        ashl    #16,r0,(r3)
-        brw     noname.7
-        nop     
-noname.18:
+;aprim=r5
+;a2=r6
+;a20=r6
+;a21=r7
+;bprim=r8
+;b2=r9
+;qprim=r10      ; initially used as q''
+;rprim=r11      ; initially used as r''
 
-        bisl2   r4,r9
 
-        movl    r9,r10
+        .psect  code,nowrt
 
-noname.3:
+.entry  bn_div_words,^m<r2,r3,r4,r5,r6,r7,r8,r9,r10,r11>
+        movl    l(ap),r2
+        movl    h(ap),r3
+        movl    d(ap),r4
+
+        movl    #0,r5
+        movl    #0,r8
+        movl    #0,r0
+;       movl    #0,r1
+
+        rotl    #-1,r2,r6       ; a20 = l >> 1 (almost)
+        rotl    #-1,r3,r7       ; a21 = h >> 1 (almost)
+        rotl    #-1,r4,r9       ; b2 = d >> 1 (almost)
+
+        tstl    r6
+        bgeq    1$
+        xorl2   #^X80000000,r6  ; fixup a20 so highest bit is 0
+        incl    r5              ; a' = 1
+1$:
+        tstl    r7
+        bgeq    2$
+        xorl2   #^X80000000,r6  ; fixup a20 so highest bit is 1,
+                                ; since that's what was lowest in a21
+        xorl2   #^X80000000,r7  ; fixup a21 so highest bit is 1
+2$:
+        tstl    r9
+        beql    666$            ; Uh-oh, the divisor is 0...
+        bgtr    3$
+        xorl2   #^X80000000,r9  ; fixup b2 so highest bit is 0
+        incl    r8              ; b' = 1
+3$:
+        tstl    r9
+        bneq    4$              ; if b2 is 0, we know that b' is 1
+        tstl    r3
+        bneq    666$            ; if higher half isn't 0, we overflow
+        movl    r2,r10          ; otherwise, we have our result
+        brb     42$             ; This is a success, really.
+4$:
+        ediv    r9,r6,r10,r11
+
+        tstl    r8
+        bneq    5$              ; If b' != 0, go to the other part
+;       addl3   r11,r11,r1
+;       addl2   r5,r1
+        brb     42$
+5$:
+        ashl    #1,r11,r11
+        subl2   r10,r11
+        addl2   r5,r11
+        bgeq    7$
+6$:
+        decl    r10
+        addl2   r4,r11
+        blss    6$
+7$:
+;       movl    r11,r1
+42$:
         movl    r10,r0
-        ret     
-        tstl    r0
-
+666$:
+        ret
 
-        .psect  code,nowrt
-
-.entry  BN_ADD_WORDS,^m<r2,r3,r4,r5,r6,r7>
-
-        tstl    16(ap)
-        bgtr    noname.21
-        clrl    r7
-        brw     noname.22
-noname.21:
-
-        clrl    r4
-
-        tstl    r0
-noname.23:
-
-        movl    8(ap),r6
-        addl3   r4,(r6),r2
-
-        bicl2   #0,r2
-
-        clrl    r0
-        cmpl    r2,r4
-        bgequ   vcg.3
-        incl    r0
-vcg.3:
-        movl    r0,r4
-
-        movl    12(ap),r5
-        addl3   (r5),r2,r1
-        bicl2   #0,r1
-
-        clrl    r0
-        cmpl    r1,r2
-        bgequ   vcg.4
-        incl    r0
-vcg.4:
-        addl2   r0,r4
-
-        movl    4(ap),r3
-        movl    r1,(r3)
-
-        decl    16(ap)
-        bgtr    gen.1
-        brw     noname.25
-gen.1:
-noname.24:
-
-        addl3   r4,4(r6),r2
-
-        bicl2   #0,r2
-
-        clrl    r0
-        cmpl    r2,r4
-        bgequ   vcg.5
-        incl    r0
-vcg.5:
-        movl    r0,r4
-
-        addl3   4(r5),r2,r1
-        bicl2   #0,r1
-
-        clrl    r0
-        cmpl    r1,r2
-        bgequ   vcg.6
-        incl    r0
-vcg.6:
-        addl2   r0,r4
-
-        movl    r1,4(r3)
-
-        decl    16(ap)
-        bleq    noname.25
-noname.26:
-
-        addl3   r4,8(r6),r2
-
-        bicl2   #0,r2
-
-        clrl    r0
-        cmpl    r2,r4
-        bgequ   vcg.7
-        incl    r0
-vcg.7:
-        movl    r0,r4
-
-        addl3   8(r5),r2,r1
-        bicl2   #0,r1
-
-        clrl    r0
-        cmpl    r1,r2
-        bgequ   vcg.8
-        incl    r0
-vcg.8:
-        addl2   r0,r4
-
-        movl    r1,8(r3)
-
-        decl    16(ap)
-        bleq    noname.25
-noname.27:
-
-        addl3   r4,12(r6),r2
-
-        bicl2   #0,r2
-
-        clrl    r0
-        cmpl    r2,r4
-        bgequ   vcg.9
-        incl    r0
-vcg.9:
-        movl    r0,r4
-
-        addl3   12(r5),r2,r1
-        bicl2   #0,r1
-
-        clrl    r0
-        cmpl    r1,r2
-        bgequ   vcg.10
-        incl    r0
-vcg.10:
-        addl2   r0,r4
+        .title  vax_bn_add_words  unsigned add of two arrays
+;
+; Richard Levitte 20-Nov-2000
+;
+; ULONG bn_add_words(ULONG r[], ULONG a[], ULONG b[], int n) {
+;       ULONG c = 0;
+;       int i;
+;       for (i = 0; i < n; i++) <c,r[i]> = a[i] + b[i] + c;
+;       return(c);
+; }
 
-        movl    r1,12(r3)
+r=4 ;(AP)       r       by reference (output)
+a=8 ;(AP)       a       by reference (input)
+b=12 ;(AP)      b       by reference (input)
+n=16 ;(AP)      n       by value (input)
 
-        decl    16(ap)
-        bleq    noname.25
-noname.28:
 
-        addl3   #16,r6,8(ap)
+        .psect  code,nowrt
 
-        addl3   #16,r5,12(ap)
+.entry  bn_add_words,^m<r2,r3,r4,r5,r6>
 
-        addl3   #16,r3,4(ap)
-        brw     noname.23
-        tstl    r0
-noname.25:
+        moval   @r(ap),r2
+        moval   @a(ap),r3
+        moval   @b(ap),r4
+        movl    n(ap),r5        ; assumed >0 by C code
+        clrl    r0              ; c
 
-        movl    r4,r7
+        tstl    r5              ; carry = 0
+        bleq    666$
 
-noname.22:
-        movl    r7,r0
-        ret     
-        nop     
+0$:
+        movl    (r3)+,r6        ; carry untouched
+        adwc    (r4)+,r6        ; carry used and touched
+        movl    r6,(r2)+        ; carry untouched
+        sobgtr  r5,0$           ; carry untouched
 
+        adwc    #0,r0
+666$:
+        ret
 
+        .title  vax_bn_sub_words  unsigned add of two arrays
+;
+; Richard Levitte 20-Nov-2000
+;
+; ULONG bn_sub_words(ULONG r[], ULONG a[], ULONG b[], int n) {
+;       ULONG c = 0;
+;       int i;
+;       for (i = 0; i < n; i++) <c,r[i]> = a[i] - b[i] - c;
+;       return(c);
+; }
 
-;r=4 ;(AP)
-;a=8 ;(AP)
-;b=12 ;(AP)
-;n=16 ;(AP)     n       by value (input)
+r=4 ;(AP)       r       by reference (output)
+a=8 ;(AP)       a       by reference (input)
+b=12 ;(AP)      b       by reference (input)
+n=16 ;(AP)      n       by value (input)
 
-        .psect  code,nowrt
 
-.entry  BN_SUB_WORDS,^m<r2,r3,r4,r5,r6,r7>
+        .psect  code,nowrt
 
-        clrl    r6
+.entry  bn_sub_words,^m<r2,r3,r4,r5,r6>
 
-        tstl    16(ap)
-        bgtr    noname.31
-        clrl    r7
-        brw     noname.32
-        tstl    r0
-noname.31:
+        moval   @r(ap),r2
+        moval   @a(ap),r3
+        moval   @b(ap),r4
+        movl    n(ap),r5        ; assumed >0 by C code
+        clrl    r0              ; c
 
-noname.33:
+        tstl    r5              ; carry = 0
+        bleq    666$
 
-        movl    8(ap),r5
-        movl    (r5),r1
-        movl    12(ap),r4
-        movl    (r4),r2
-
-        movl    4(ap),r3
-        subl3   r2,r1,r0
-        subl2   r6,r0
-        bicl3   #0,r0,(r3)
-
-        cmpl    r1,r2
-        beql    noname.34
-        clrl    r0
-        cmpl    r1,r2
-        bgequ   vcg.11
-        incl    r0
-vcg.11:
-        movl    r0,r6
-noname.34:
-
-        decl    16(ap)
-        bgtr    gen.2
-        brw     noname.36
-gen.2:
-noname.35:
-
-        movl    4(r5),r2
-        movl    4(r4),r1
-
-        subl3   r1,r2,r0
-        subl2   r6,r0
-        bicl3   #0,r0,4(r3)
-
-        cmpl    r2,r1
-        beql    noname.37
-        clrl    r0
-        cmpl    r2,r1
-        bgequ   vcg.12
-        incl    r0
-vcg.12:
-        movl    r0,r6
-noname.37:
-
-        decl    16(ap)
-        bleq    noname.36
-noname.38:
-
-        movl    8(r5),r1
-        movl    8(r4),r2
-
-        subl3   r2,r1,r0
-        subl2   r6,r0
-        bicl3   #0,r0,8(r3)
-
-        cmpl    r1,r2
-        beql    noname.39
-        clrl    r0
-        cmpl    r1,r2
-        bgequ   vcg.13
-        incl    r0
-vcg.13:
-        movl    r0,r6
-noname.39:
-
-        decl    16(ap)
-        bleq    noname.36
-noname.40:
-
-        movl    12(r5),r1
-        movl    12(r4),r2
-
-        subl3   r2,r1,r0
-        subl2   r6,r0
-        bicl3   #0,r0,12(r3)
-
-        cmpl    r1,r2
-        beql    noname.41
-        clrl    r0
-        cmpl    r1,r2
-        bgequ   vcg.14
-        incl    r0
-vcg.14:
-        movl    r0,r6
-noname.41:
-
-        decl    16(ap)
-        bleq    noname.36
-noname.42:
-
-        addl3   #16,r5,8(ap)
-
-        addl3   #16,r4,12(ap)
-
-        addl3   #16,r3,4(ap)
-        brw     noname.33
-        tstl    r0
-noname.36:
-
-        movl    r6,r7
-
-noname.32:
-        movl    r7,r0
-        ret     
-        nop     
+0$:
+        movl    (r3)+,r6        ; carry untouched
+        sbwc    (r4)+,r6        ; carry used and touched
+        movl    r6,(r2)+        ; carry untouched
+        sobgtr  r5,0$           ; carry untouched
 
+        adwc    #0,r0
+666$:
+        ret
 
 
 ;r=4 ;(AP)
@@ -6615,81 +6410,3 @@ noname.610:
 
 ; For now, the code below doesn't work, so I end this prematurely.
 .end
-
-        .title  vax_bn_div64    division 64/32=>32
-; 
-; r.l. 16-jan-1998
-;
-; unsigned int bn_div64(unsigned long h, unsigned long l, unsigned long d)
-;       return <h,l>/d;
-;
-
-        .psect  code,nowrt
-
-h=4 ;(AP)       by value (input)
-l=8 ;(AP)       by value (input)
-d=12 ;(AP)      by value (input)
-
-.entry  bn_div64,^m<r2,r3,r4,r5,r6,r7,r8,r9>
-
-        movl    l(ap),r2        ; l
-        movl    h(ap),r3        ; h
-        movl    d(ap),r4        ; d
-        clrl    r5              ; q
-        clrl    r6              ; r
-
-        ; Treat "negative" specially
-        tstl    r3
-        blss    30$
-
-        tstl    r4
-        beql    90$
-
-        ediv    r4,r2,r5,r6
-        bvs     666$
-
-        movl    r5,r0
-        ret
-
-30$:
-        ; The theory here is to do some harmless shifting and a little
-        ; bit of rounding (brackets are to designate when decimals are
-        ; cut off):
-        ;
-        ;       result = 2 * [ ([<h,0>/2] + [d/2]) / d ] + [ l / d ]
-
-        movl    #0,r7
-        movl    r3,r8           ; copy h
-        ashq    #-1,r7,r7       ; [<h,0>/2] => <r8,r7>
-        bicl2   #^X80000000,r8  ; Remove "sign"
-
-        movl    r4,r9           ; copy d
-        ashl    #-1,r9,r9       ; [d/2] => r9
-        bicl2   #^X80000000,r9  ; Remove "sign"
-
-        addl2   r9,r7
-        adwc    #0,r8           ; [<h,0>/2] + [d/2] => <r8,r7>
-
-        ediv    r4,r7,r5,r6     ; [ ([<h,0>/2] + [d/2]) / d ] => <r5,r6>
-        bvs     666$
-
-        movl    #0,r6
-        ashq    #1,r5,r5        ; 2 * [ ([<h,0>/2] + [d/2]) / d ] => r5
-
-        movl    #0,r3
-        ediv    r4,r2,r8,r9     ; [ l / d ] => <r8,r9>
-
-        addl2   r8,r5           ;
-        bcs     666$
-
-        movl    r5,r0
-        ret
-                
-90$:
-        movl    #-1,r0
-        ret
-
-666$:
-
-        
-.end
diff --git a/src/lib/libcrypto/bn/bnspeed.c b/src/lib/libcrypto/bn/bnspeed.c
index f7c2790fff..b554ac8cf8 100644
--- a/src/lib/libcrypto/bn/bnspeed.c
+++ b/src/lib/libcrypto/bn/bnspeed.c
@@ -1,3 +1,5 @@
+/* unused */
+
 /* crypto/bn/bnspeed.c */
 /* Copyright (C) 1995-1998 Eric Young (eay@cryptsoft.com)
  * All rights reserved.
@@ -66,14 +68,13 @@
 #include <stdlib.h>
 #include <signal.h>
 #include <string.h>
-#include "crypto.h"
-#include "err.h"
+#include <openssl/crypto.h>
+#include <openssl/err.h>
 
-#ifndef MSDOS
+#if !defined(OPENSSL_SYS_MSDOS) && (!defined(OPENSSL_SYS_VMS) || defined(__DECC)) && !defined(OPENSSL_SYS_MACOSX)
 #define TIMES
 #endif
 
-#ifndef VMS
 #ifndef _IRIX
 #include <time.h>
 #endif
@@ -81,36 +82,33 @@
 #include <sys/types.h>
 #include <sys/times.h>
 #endif
-#else /* VMS */
-#include <types.h>
-struct tms {
-        time_t tms_utime;
-        time_t tms_stime;
-        time_t tms_uchild;      /* I dunno...  */
-        time_t tms_uchildsys;   /* so these names are a guess :-) */
-        }
+
+/* Depending on the VMS version, the tms structure is perhaps defined.
+   The __TMS macro will show if it was.  If it wasn't defined, we should
+   undefine TIMES, since that tells the rest of the program how things
+   should be handled.                           -- Richard Levitte */
+#if defined(OPENSSL_SYS_VMS_DECC) && !defined(__TMS)
+#undef TIMES
 #endif
+
 #ifndef TIMES
 #include <sys/timeb.h>
 #endif
 
-#ifdef sun
+#if defined(sun) || defined(__ultrix)
+#define _POSIX_SOURCE
 #include <limits.h>
 #include <sys/param.h>
 #endif
 
-#include "bn.h"
-#include "x509.h"
+#include <openssl/bn.h>
+#include <openssl/x509.h>
 
 /* The following if from times(3) man page.  It may need to be changed */
 #ifndef HZ
 # ifndef CLK_TCK
 #  ifndef _BSD_CLK_TCK_ /* FreeBSD hack */
-#   ifndef VMS
-#    define HZ  100.0
-#   else /* VMS */
-#    define HZ  100.0
-#   endif
+#   define HZ   100.0
 #  else /* _BSD_CLK_TCK_ */
 #   define HZ ((double)_BSD_CLK_TCK_)
 #  endif
@@ -123,17 +121,11 @@ struct tms {
 #define BUFSIZE ((long)1024*8)
 int run=0;
 
-#ifndef NOPROTO
 static double Time_F(int s);
-#else
-static double Time_F();
-#endif
-
 #define START   0
 #define STOP    1
 
-static double Time_F(s)
-int s;
+static double Time_F(int s)
         {
         double ret;
 #ifdef TIMES
@@ -175,27 +167,20 @@ static int sizes[NUM_SIZES]={128,256,512,1024,2048};
 
 void do_mul(BIGNUM *r,BIGNUM *a,BIGNUM *b,BN_CTX *ctx); 
 
-int main(argc,argv)
-int argc;
-char **argv;
+int main(int argc, char **argv)
         {
         BN_CTX *ctx;
-        BIGNUM *a,*b,*c,*r;
+        BIGNUM a,b,c;
 
         ctx=BN_CTX_new();
-        a=BN_new();
-        b=BN_new();
-        c=BN_new();
-        r=BN_new();
+        BN_init(&a);
+        BN_init(&b);
+        BN_init(&c);
 
-        do_mul(a,b,c,ctx);
+        do_mul(&a,&b,&c,ctx);
         }
 
-void do_mul(r,a,b,ctx)
-BIGNUM *r;
-BIGNUM *a;
-BIGNUM *b;
-BN_CTX *ctx;
+void do_mul(BIGNUM *r, BIGNUM *a, BIGNUM *b, BN_CTX *ctx)
         {
         int i,j,k;
         double tm;
@@ -211,7 +196,7 @@ BN_CTX *ctx;
                         BN_rand(b,sizes[j],1,0);
                         Time_F(START);
                         for (k=0; k<num; k++)
-                                BN_mul(r,b,a);
+                                BN_mul(r,b,a,ctx);
                         tm=Time_F(STOP);
                         printf("mul %4d x %4d -> %8.3fms\n",sizes[i],sizes[j],tm*1000.0/num);
                         }
diff --git a/src/lib/libcrypto/bn/bntest.c b/src/lib/libcrypto/bn/bntest.c
index 9ebd68b429..443cf420e5 100644
--- a/src/lib/libcrypto/bn/bntest.c
+++ b/src/lib/libcrypto/bn/bntest.c
@@ -59,65 +59,70 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+
 #include "e_os.h"
-#include "bio.h"
-#include "bn.h"
-#include "rand.h"
-#include "x509.h"
-#include "err.h"
 
-#ifdef WINDOWS
+#include <openssl/bio.h>
+#include <openssl/bn.h>
+#include <openssl/rand.h>
+#include <openssl/x509.h>
+#include <openssl/err.h>
+
+#ifdef OPENSSL_SYS_WINDOWS
 #include "../bio/bss_file.c"
 #endif
 
-#ifndef NOPROTO
-int test_add (BIO *bp);
-int test_sub (BIO *bp);
-int test_lshift1 (BIO *bp);
-int test_lshift (BIO *bp);
-int test_rshift1 (BIO *bp);
-int test_rshift (BIO *bp);
-int test_div (BIO *bp,BN_CTX *ctx);
-int test_mul (BIO *bp);
-int test_sqr (BIO *bp,BN_CTX *ctx);
-int test_mont (BIO *bp,BN_CTX *ctx);
-int test_mod (BIO *bp,BN_CTX *ctx);
-int test_mod_mul (BIO *bp,BN_CTX *ctx);
-int test_mod_exp (BIO *bp,BN_CTX *ctx);
+const int num0 = 100; /* number of tests */
+const int num1 = 50;  /* additional tests for some functions */
+const int num2 = 5;   /* number of tests for slow functions */
+
+int test_add(BIO *bp);
+int test_sub(BIO *bp);
+int test_lshift1(BIO *bp);
+int test_lshift(BIO *bp,BN_CTX *ctx,BIGNUM *a_);
+int test_rshift1(BIO *bp);
+int test_rshift(BIO *bp,BN_CTX *ctx);
+int test_div(BIO *bp,BN_CTX *ctx);
+int test_div_recp(BIO *bp,BN_CTX *ctx);
+int test_mul(BIO *bp);
+int test_sqr(BIO *bp,BN_CTX *ctx);
+int test_mont(BIO *bp,BN_CTX *ctx);
+int test_mod(BIO *bp,BN_CTX *ctx);
+int test_mod_mul(BIO *bp,BN_CTX *ctx);
+int test_mod_exp(BIO *bp,BN_CTX *ctx);
+int test_exp(BIO *bp,BN_CTX *ctx);
+int test_kron(BIO *bp,BN_CTX *ctx);
+int test_sqrt(BIO *bp,BN_CTX *ctx);
 int rand_neg(void);
-#else
-int test_add ();
-int test_sub ();
-int test_lshift1 ();
-int test_lshift ();
-int test_rshift1 ();
-int test_rshift ();
-int test_div ();
-int test_mul ();
-int test_sqr ();
-int test_mont ();
-int test_mod ();
-int test_mod_mul ();
-int test_mod_exp ();
-int rand_neg();
-#endif
-
 static int results=0;
 
-#ifdef NO_STDIO
+#ifdef OPENSSL_NO_STDIO
 #define APPS_WIN16
 #include "bss_file.c"
 #endif
 
-int main(argc,argv)
-int argc;
-char *argv[];
+static unsigned char lst[]="\xC6\x4F\x43\x04\x2A\xEA\xCA\x6E\x58\x36\x80\x5B\xE8\xC9"
+"\x9B\x04\x5D\x48\x36\xC2\xFD\x16\xC9\x64\xF0";
+
+static const char rnd_seed[] = "string to make the random number generator think it has entropy";
+
+static void message(BIO *out, char *m)
+        {
+        fprintf(stderr, "test %s\n", m);
+        BIO_puts(out, "print \"test ");
+        BIO_puts(out, m);
+        BIO_puts(out, "\\n\"\n");
+        }
+
+int main(int argc, char *argv[])
         {
         BN_CTX *ctx;
         BIO *out;
         char *outfile=NULL;
 
-        srand((unsigned int)time(NULL));
+        results = 0;
+
+        RAND_seed(rnd_seed, sizeof rnd_seed); /* or BN_generate_prime may fail */
 
         argc--;
         argv++;
@@ -156,361 +161,516 @@ char *argv[];
         if (!results)
                 BIO_puts(out,"obase=16\nibase=16\n");
 
-        fprintf(stderr,"test BN_add\n");
+        message(out,"BN_add");
         if (!test_add(out)) goto err;
-        fflush(stdout);
+        BIO_flush(out);
 
-        fprintf(stderr,"test BN_sub\n");
+        message(out,"BN_sub");
         if (!test_sub(out)) goto err;
-        fflush(stdout);
+        BIO_flush(out);
 
-        fprintf(stderr,"test BN_lshift1\n");
+        message(out,"BN_lshift1");
         if (!test_lshift1(out)) goto err;
-        fflush(stdout);
+        BIO_flush(out);
+
+        message(out,"BN_lshift (fixed)");
+        if (!test_lshift(out,ctx,BN_bin2bn(lst,sizeof(lst)-1,NULL)))
+            goto err;
+        BIO_flush(out);
 
-        fprintf(stderr,"test BN_lshift\n");
-        if (!test_lshift(out)) goto err;
-        fflush(stdout);
+        message(out,"BN_lshift");
+        if (!test_lshift(out,ctx,NULL)) goto err;
+        BIO_flush(out);
 
-        fprintf(stderr,"test BN_rshift1\n");
+        message(out,"BN_rshift1");
         if (!test_rshift1(out)) goto err;
-        fflush(stdout);
+        BIO_flush(out);
 
-        fprintf(stderr,"test BN_rshift\n");
-        if (!test_rshift(out)) goto err;
-        fflush(stdout);
+        message(out,"BN_rshift");
+        if (!test_rshift(out,ctx)) goto err;
+        BIO_flush(out);
 
-        fprintf(stderr,"test BN_sqr\n");
+        message(out,"BN_sqr");
         if (!test_sqr(out,ctx)) goto err;
-        fflush(stdout);
+        BIO_flush(out);
 
-        fprintf(stderr,"test BN_mul\n");
+        message(out,"BN_mul");
         if (!test_mul(out)) goto err;
-        fflush(stdout);
+        BIO_flush(out);
 
-        fprintf(stderr,"test BN_div\n");
+        message(out,"BN_div");
         if (!test_div(out,ctx)) goto err;
-        fflush(stdout);
+        BIO_flush(out);
 
-        fprintf(stderr,"test BN_mod\n");
+        message(out,"BN_div_recp");
+        if (!test_div_recp(out,ctx)) goto err;
+        BIO_flush(out);
+
+        message(out,"BN_mod");
         if (!test_mod(out,ctx)) goto err;
-        fflush(stdout);
+        BIO_flush(out);
 
-        fprintf(stderr,"test BN_mod_mul\n");
+        message(out,"BN_mod_mul");
         if (!test_mod_mul(out,ctx)) goto err;
-        fflush(stdout);
+        BIO_flush(out);
 
-/*
-        fprintf(stderr,"test BN_mont\n");
+        message(out,"BN_mont");
         if (!test_mont(out,ctx)) goto err;
-        fflush(stdout);
-*/
-        fprintf(stderr,"test BN_mod_exp\n");
+        BIO_flush(out);
+
+        message(out,"BN_mod_exp");
         if (!test_mod_exp(out,ctx)) goto err;
-        fflush(stdout);
+        BIO_flush(out);
+
+        message(out,"BN_exp");
+        if (!test_exp(out,ctx)) goto err;
+        BIO_flush(out);
+
+        message(out,"BN_kronecker");
+        if (!test_kron(out,ctx)) goto err;
+        BIO_flush(out);
+
+        message(out,"BN_mod_sqrt");
+        if (!test_sqrt(out,ctx)) goto err;
+        BIO_flush(out);
+
+        BN_CTX_free(ctx);
+        BIO_free(out);
 
 /**/
         exit(0);
 err:
+        BIO_puts(out,"1\n"); /* make sure the Perl script fed by bc notices
+                              * the failure, see test_bn in test/Makefile.ssl*/
+        BIO_flush(out);
         ERR_load_crypto_strings();
-        ERR_print_errors(out);
+        ERR_print_errors_fp(stderr);
         exit(1);
         return(1);
         }
 
-int test_add(bp)
-BIO *bp;
+int test_add(BIO *bp)
         {
-        BIGNUM *a,*b,*c;
+        BIGNUM a,b,c;
         int i;
-        int j;
 
-        a=BN_new();
-        b=BN_new();
-        c=BN_new();
+        BN_init(&a);
+        BN_init(&b);
+        BN_init(&c);
 
-        BN_rand(a,512,0,0);
-        for (i=0; i<100; i++)
+        BN_bntest_rand(&a,512,0,0);
+        for (i=0; i<num0; i++)
                 {
-                BN_rand(b,450+i,0,0);
-                a->neg=rand_neg();
-                b->neg=rand_neg();
-                if (bp == NULL)
-                        for (j=0; j<10000; j++)
-                                BN_add(c,a,b);
-                BN_add(c,a,b);
+                BN_bntest_rand(&b,450+i,0,0);
+                a.neg=rand_neg();
+                b.neg=rand_neg();
+                BN_add(&c,&a,&b);
                 if (bp != NULL)
                         {
                         if (!results)
                                 {
-                                BN_print(bp,a);
+                                BN_print(bp,&a);
                                 BIO_puts(bp," + ");
-                                BN_print(bp,b);
+                                BN_print(bp,&b);
                                 BIO_puts(bp," - ");
                                 }
-                        BN_print(bp,c);
+                        BN_print(bp,&c);
                         BIO_puts(bp,"\n");
                         }
+                a.neg=!a.neg;
+                b.neg=!b.neg;
+                BN_add(&c,&c,&b);
+                BN_add(&c,&c,&a);
+                if(!BN_is_zero(&c))
+                    {
+                    fprintf(stderr,"Add test failed!\n");
+                    return 0;
+                    }
                 }
-        BN_free(a);
-        BN_free(b);
-        BN_free(c);
+        BN_free(&a);
+        BN_free(&b);
+        BN_free(&c);
         return(1);
         }
 
-int test_sub(bp)
-BIO *bp;
+int test_sub(BIO *bp)
         {
-        BIGNUM *a,*b,*c;
+        BIGNUM a,b,c;
         int i;
-        int j;
 
-        a=BN_new();
-        b=BN_new();
-        c=BN_new();
+        BN_init(&a);
+        BN_init(&b);
+        BN_init(&c);
 
-        BN_rand(a,512,0,0);
-        for (i=0; i<100; i++)
+        for (i=0; i<num0+num1; i++)
                 {
-                BN_rand(b,400+i,0,0);
-                a->neg=rand_neg();
-                b->neg=rand_neg();
-                if (bp == NULL)
-                        for (j=0; j<10000; j++)
-                                BN_sub(c,a,b);
-                BN_sub(c,a,b);
+                if (i < num1)
+                        {
+                        BN_bntest_rand(&a,512,0,0);
+                        BN_copy(&b,&a);
+                        if (BN_set_bit(&a,i)==0) return(0);
+                        BN_add_word(&b,i);
+                        }
+                else
+                        {
+                        BN_bntest_rand(&b,400+i-num1,0,0);
+                        a.neg=rand_neg();
+                        b.neg=rand_neg();
+                        }
+                BN_sub(&c,&a,&b);
                 if (bp != NULL)
                         {
                         if (!results)
                                 {
-                                BN_print(bp,a);
+                                BN_print(bp,&a);
                                 BIO_puts(bp," - ");
-                                BN_print(bp,b);
+                                BN_print(bp,&b);
                                 BIO_puts(bp," - ");
                                 }
-                        BN_print(bp,c);
+                        BN_print(bp,&c);
                         BIO_puts(bp,"\n");
                         }
+                BN_add(&c,&c,&b);
+                BN_sub(&c,&c,&a);
+                if(!BN_is_zero(&c))
+                    {
+                    fprintf(stderr,"Subtract test failed!\n");
+                    return 0;
+                    }
                 }
-        BN_free(a);
-        BN_free(b);
-        BN_free(c);
+        BN_free(&a);
+        BN_free(&b);
+        BN_free(&c);
         return(1);
         }
 
-int test_div(bp,ctx)
-BIO *bp;
-BN_CTX *ctx;
+int test_div(BIO *bp, BN_CTX *ctx)
         {
-        BIGNUM *a,*b,*c,*d;
+        BIGNUM a,b,c,d,e;
         int i;
-        int j;
 
-        a=BN_new();
-        b=BN_new();
-        c=BN_new();
-        d=BN_new();
+        BN_init(&a);
+        BN_init(&b);
+        BN_init(&c);
+        BN_init(&d);
+        BN_init(&e);
 
-        BN_rand(a,400,0,0);
-        for (i=0; i<100; i++)
+        for (i=0; i<num0+num1; i++)
                 {
-                BN_rand(b,50+i,0,0);
-                a->neg=rand_neg();
-                b->neg=rand_neg();
-                if (bp == NULL)
-                        for (j=0; j<100; j++)
-                                BN_div(d,c,a,b,ctx);
-                BN_div(d,c,a,b,ctx);
+                if (i < num1)
+                        {
+                        BN_bntest_rand(&a,400,0,0);
+                        BN_copy(&b,&a);
+                        BN_lshift(&a,&a,i);
+                        BN_add_word(&a,i);
+                        }
+                else
+                        BN_bntest_rand(&b,50+3*(i-num1),0,0);
+                a.neg=rand_neg();
+                b.neg=rand_neg();
+                BN_div(&d,&c,&a,&b,ctx);
                 if (bp != NULL)
                         {
                         if (!results)
                                 {
-                                BN_print(bp,a);
+                                BN_print(bp,&a);
                                 BIO_puts(bp," / ");
-                                BN_print(bp,b);
+                                BN_print(bp,&b);
                                 BIO_puts(bp," - ");
                                 }
-                        BN_print(bp,d);
+                        BN_print(bp,&d);
                         BIO_puts(bp,"\n");
 
                         if (!results)
                                 {
-                                BN_print(bp,a);
+                                BN_print(bp,&a);
                                 BIO_puts(bp," % ");
-                                BN_print(bp,b);
+                                BN_print(bp,&b);
                                 BIO_puts(bp," - ");
                                 }
-                        BN_print(bp,c);
+                        BN_print(bp,&c);
                         BIO_puts(bp,"\n");
                         }
+                BN_mul(&e,&d,&b,ctx);
+                BN_add(&d,&e,&c);
+                BN_sub(&d,&d,&a);
+                if(!BN_is_zero(&d))
+                    {
+                    fprintf(stderr,"Division test failed!\n");
+                    return 0;
+                    }
                 }
-        BN_free(a);
-        BN_free(b);
-        BN_free(c);
-        BN_free(d);
+        BN_free(&a);
+        BN_free(&b);
+        BN_free(&c);
+        BN_free(&d);
+        BN_free(&e);
         return(1);
         }
 
-int test_mul(bp)
-BIO *bp;
+int test_div_recp(BIO *bp, BN_CTX *ctx)
         {
-        BIGNUM *a,*b,*c;
+        BIGNUM a,b,c,d,e;
+        BN_RECP_CTX recp;
         int i;
-        int j;
 
-        a=BN_new();
-        b=BN_new();
-        c=BN_new();
+        BN_RECP_CTX_init(&recp);
+        BN_init(&a);
+        BN_init(&b);
+        BN_init(&c);
+        BN_init(&d);
+        BN_init(&e);
 
-        BN_rand(a,200,0,0);
-        for (i=0; i<100; i++)
+        for (i=0; i<num0+num1; i++)
                 {
-                BN_rand(b,250+i,0,0);
-                a->neg=rand_neg();
-                b->neg=rand_neg();
-                if (bp == NULL)
-                        for (j=0; j<100; j++)
-                                BN_mul(c,a,b);
-                BN_mul(c,a,b);
+                if (i < num1)
+                        {
+                        BN_bntest_rand(&a,400,0,0);
+                        BN_copy(&b,&a);
+                        BN_lshift(&a,&a,i);
+                        BN_add_word(&a,i);
+                        }
+                else
+                        BN_bntest_rand(&b,50+3*(i-num1),0,0);
+                a.neg=rand_neg();
+                b.neg=rand_neg();
+                BN_RECP_CTX_set(&recp,&b,ctx);
+                BN_div_recp(&d,&c,&a,&recp,ctx);
                 if (bp != NULL)
                         {
                         if (!results)
                                 {
-                                BN_print(bp,a);
+                                BN_print(bp,&a);
+                                BIO_puts(bp," / ");
+                                BN_print(bp,&b);
+                                BIO_puts(bp," - ");
+                                }
+                        BN_print(bp,&d);
+                        BIO_puts(bp,"\n");
+
+                        if (!results)
+                                {
+                                BN_print(bp,&a);
+                                BIO_puts(bp," % ");
+                                BN_print(bp,&b);
+                                BIO_puts(bp," - ");
+                                }
+                        BN_print(bp,&c);
+                        BIO_puts(bp,"\n");
+                        }
+                BN_mul(&e,&d,&b,ctx);
+                BN_add(&d,&e,&c);
+                BN_sub(&d,&d,&a);
+                if(!BN_is_zero(&d))
+                    {
+                    fprintf(stderr,"Reciprocal division test failed!\n");
+                    fprintf(stderr,"a=");
+                    BN_print_fp(stderr,&a);
+                    fprintf(stderr,"\nb=");
+                    BN_print_fp(stderr,&b);
+                    fprintf(stderr,"\n");
+                    return 0;
+                    }
+                }
+        BN_free(&a);
+        BN_free(&b);
+        BN_free(&c);
+        BN_free(&d);
+        BN_free(&e);
+        BN_RECP_CTX_free(&recp);
+        return(1);
+        }
+
+int test_mul(BIO *bp)
+        {
+        BIGNUM a,b,c,d,e;
+        int i;
+        BN_CTX *ctx;
+
+        ctx = BN_CTX_new();
+        if (ctx == NULL) exit(1);
+        
+        BN_init(&a);
+        BN_init(&b);
+        BN_init(&c);
+        BN_init(&d);
+        BN_init(&e);
+
+        for (i=0; i<num0+num1; i++)
+                {
+                if (i <= num1)
+                        {
+                        BN_bntest_rand(&a,100,0,0);
+                        BN_bntest_rand(&b,100,0,0);
+                        }
+                else
+                        BN_bntest_rand(&b,i-num1,0,0);
+                a.neg=rand_neg();
+                b.neg=rand_neg();
+                BN_mul(&c,&a,&b,ctx);
+                if (bp != NULL)
+                        {
+                        if (!results)
+                                {
+                                BN_print(bp,&a);
                                 BIO_puts(bp," * ");
-                                BN_print(bp,b);
+                                BN_print(bp,&b);
                                 BIO_puts(bp," - ");
                                 }
-                        BN_print(bp,c);
+                        BN_print(bp,&c);
                         BIO_puts(bp,"\n");
                         }
+                BN_div(&d,&e,&c,&a,ctx);
+                BN_sub(&d,&d,&b);
+                if(!BN_is_zero(&d) || !BN_is_zero(&e))
+                    {
+                    fprintf(stderr,"Multiplication test failed!\n");
+                    return 0;
+                    }
                 }
-        BN_free(a);
-        BN_free(b);
-        BN_free(c);
+        BN_free(&a);
+        BN_free(&b);
+        BN_free(&c);
+        BN_free(&d);
+        BN_free(&e);
+        BN_CTX_free(ctx);
         return(1);
         }
 
-int test_sqr(bp,ctx)
-BIO *bp;
-BN_CTX *ctx;
+int test_sqr(BIO *bp, BN_CTX *ctx)
         {
-        BIGNUM *a,*c;
+        BIGNUM a,c,d,e;
         int i;
-        int j;
 
-        a=BN_new();
-        c=BN_new();
+        BN_init(&a);
+        BN_init(&c);
+        BN_init(&d);
+        BN_init(&e);
 
-        for (i=0; i<40; i++)
+        for (i=0; i<num0; i++)
                 {
-                BN_rand(a,40+i*10,0,0);
-                a->neg=rand_neg();
-                if (bp == NULL)
-                        for (j=0; j<100; j++)
-                                BN_sqr(c,a,ctx);
-                BN_sqr(c,a,ctx);
+                BN_bntest_rand(&a,40+i*10,0,0);
+                a.neg=rand_neg();
+                BN_sqr(&c,&a,ctx);
                 if (bp != NULL)
                         {
                         if (!results)
                                 {
-                                BN_print(bp,a);
+                                BN_print(bp,&a);
                                 BIO_puts(bp," * ");
-                                BN_print(bp,a);
+                                BN_print(bp,&a);
                                 BIO_puts(bp," - ");
                                 }
-                        BN_print(bp,c);
+                        BN_print(bp,&c);
                         BIO_puts(bp,"\n");
                         }
+                BN_div(&d,&e,&c,&a,ctx);
+                BN_sub(&d,&d,&a);
+                if(!BN_is_zero(&d) || !BN_is_zero(&e))
+                    {
+                    fprintf(stderr,"Square test failed!\n");
+                    return 0;
+                    }
                 }
-        BN_free(a);
-        BN_free(c);
+        BN_free(&a);
+        BN_free(&c);
+        BN_free(&d);
+        BN_free(&e);
         return(1);
         }
 
-int test_mont(bp,ctx)
-BIO *bp;
-BN_CTX *ctx;
+int test_mont(BIO *bp, BN_CTX *ctx)
         {
-        BIGNUM *a,*b,*c,*A,*B;
-        BIGNUM *n;
+        BIGNUM a,b,c,d,A,B;
+        BIGNUM n;
         int i;
-        int j;
         BN_MONT_CTX *mont;
 
-        a=BN_new();
-        b=BN_new();
-        c=BN_new();
-        A=BN_new();
-        B=BN_new();
-        n=BN_new();
+        BN_init(&a);
+        BN_init(&b);
+        BN_init(&c);
+        BN_init(&d);
+        BN_init(&A);
+        BN_init(&B);
+        BN_init(&n);
 
         mont=BN_MONT_CTX_new();
 
-        BN_rand(a,100,0,0); /**/
-        BN_rand(b,100,0,0); /**/
-        for (i=0; i<10; i++)
+        BN_bntest_rand(&a,100,0,0); /**/
+        BN_bntest_rand(&b,100,0,0); /**/
+        for (i=0; i<num2; i++)
                 {
-                BN_rand(n,(100%BN_BITS2+1)*BN_BITS2*i*BN_BITS2,0,1); /**/
-                BN_MONT_CTX_set(mont,n,ctx);
+                int bits = (200*(i+1))/num2;
+
+                if (bits == 0)
+                        continue;
+                BN_bntest_rand(&n,bits,0,1);
+                BN_MONT_CTX_set(mont,&n,ctx);
 
-                BN_to_montgomery(A,a,mont,ctx);
-                BN_to_montgomery(B,b,mont,ctx);
+                BN_nnmod(&a,&a,&n,ctx);
+                BN_nnmod(&b,&b,&n,ctx);
 
-                if (bp == NULL)
-                        for (j=0; j<100; j++)
-                                BN_mod_mul_montgomery(c,A,B,mont,ctx);/**/
-                BN_mod_mul_montgomery(c,A,B,mont,ctx);/**/
-                BN_from_montgomery(A,c,mont,ctx);/**/
+                BN_to_montgomery(&A,&a,mont,ctx);
+                BN_to_montgomery(&B,&b,mont,ctx);
+
+                BN_mod_mul_montgomery(&c,&A,&B,mont,ctx);/**/
+                BN_from_montgomery(&A,&c,mont,ctx);/**/
                 if (bp != NULL)
                         {
                         if (!results)
                                 {
 #ifdef undef
 fprintf(stderr,"%d * %d %% %d\n",
-BN_num_bits(a),
-BN_num_bits(b),
+BN_num_bits(&a),
+BN_num_bits(&b),
 BN_num_bits(mont->N));
 #endif
-                                BN_print(bp,a);
+                                BN_print(bp,&a);
                                 BIO_puts(bp," * ");
-                                BN_print(bp,b);
+                                BN_print(bp,&b);
                                 BIO_puts(bp," % ");
-                                BN_print(bp,mont->N);
+                                BN_print(bp,&(mont->N));
                                 BIO_puts(bp," - ");
                                 }
-                        BN_print(bp,A);
+                        BN_print(bp,&A);
                         BIO_puts(bp,"\n");
                         }
+                BN_mod_mul(&d,&a,&b,&n,ctx);
+                BN_sub(&d,&d,&A);
+                if(!BN_is_zero(&d))
+                    {
+                    fprintf(stderr,"Montgomery multiplication test failed!\n");
+                    return 0;
+                    }
                 }
         BN_MONT_CTX_free(mont);
-        BN_free(a);
-        BN_free(b);
-        BN_free(c);
+        BN_free(&a);
+        BN_free(&b);
+        BN_free(&c);
+        BN_free(&d);
+        BN_free(&A);
+        BN_free(&B);
+        BN_free(&n);
         return(1);
         }
 
-int test_mod(bp,ctx)
-BIO *bp;
-BN_CTX *ctx;
+int test_mod(BIO *bp, BN_CTX *ctx)
         {
-        BIGNUM *a,*b,*c;
+        BIGNUM *a,*b,*c,*d,*e;
         int i;
-        int j;
 
         a=BN_new();
         b=BN_new();
         c=BN_new();
+        d=BN_new();
+        e=BN_new();
 
-        BN_rand(a,1024,0,0); /**/
-        for (i=0; i<20; i++)
+        BN_bntest_rand(a,1024,0,0); /**/
+        for (i=0; i<num0; i++)
                 {
-                BN_rand(b,450+i*10,0,0); /**/
+                BN_bntest_rand(b,450+i*10,0,0); /**/
                 a->neg=rand_neg();
                 b->neg=rand_neg();
-                if (bp == NULL)
-                        for (j=0; j<100; j++)
-                                BN_mod(c,a,b,ctx);/**/
                 BN_mod(c,a,b,ctx);/**/
                 if (bp != NULL)
                         {
@@ -524,19 +684,26 @@ BN_CTX *ctx;
                         BN_print(bp,c);
                         BIO_puts(bp,"\n");
                         }
+                BN_div(d,e,a,b,ctx);
+                BN_sub(e,e,c);
+                if(!BN_is_zero(e))
+                    {
+                    fprintf(stderr,"Modulo test failed!\n");
+                    return 0;
+                    }
                 }
         BN_free(a);
         BN_free(b);
         BN_free(c);
+        BN_free(d);
+        BN_free(e);
         return(1);
         }
 
-int test_mod_mul(bp,ctx)
-BIO *bp;
-BN_CTX *ctx;
+int test_mod_mul(BIO *bp, BN_CTX *ctx)
         {
         BIGNUM *a,*b,*c,*d,*e;
-        int i;
+        int i,j;
 
         a=BN_new();
         b=BN_new();
@@ -544,17 +711,14 @@ BN_CTX *ctx;
         d=BN_new();
         e=BN_new();
 
-        BN_rand(c,1024,0,0); /**/
-        for (i=0; i<10; i++)
+        for (j=0; j<3; j++) {
+        BN_bntest_rand(c,1024,0,0); /**/
+        for (i=0; i<num0; i++)
                 {
-                BN_rand(a,475+i*10,0,0); /**/
-                BN_rand(b,425+i*10,0,0); /**/
+                BN_bntest_rand(a,475+i*10,0,0); /**/
+                BN_bntest_rand(b,425+i*11,0,0); /**/
                 a->neg=rand_neg();
                 b->neg=rand_neg();
-        /*      if (bp == NULL)
-                        for (j=0; j<100; j++)
-                                BN_mod_mul(d,a,b,c,ctx);*/ /**/
-
                 if (!BN_mod_mul(e,a,b,c,ctx))
                         {
                         unsigned long l;
@@ -573,12 +737,32 @@ BN_CTX *ctx;
                                 BN_print(bp,b);
                                 BIO_puts(bp," % ");
                                 BN_print(bp,c);
+                                if ((a->neg ^ b->neg) && !BN_is_zero(e))
+                                        {
+                                        /* If  (a*b) % c  is negative,  c  must be added
+                                         * in order to obtain the normalized remainder
+                                         * (new with OpenSSL 0.9.7, previous versions of
+                                         * BN_mod_mul could generate negative results)
+                                         */
+                                        BIO_puts(bp," + ");
+                                        BN_print(bp,c);
+                                        }
                                 BIO_puts(bp," - ");
                                 }
                         BN_print(bp,e);
                         BIO_puts(bp,"\n");
                         }
+                BN_mul(d,a,b,ctx);
+                BN_sub(d,d,e);
+                BN_div(a,b,d,c,ctx);
+                if(!BN_is_zero(b))
+                    {
+                    fprintf(stderr,"Modulo multiply test failed!\n");
+                    ERR_print_errors_fp(stderr);
+                    return 0;
+                    }
                 }
+        }
         BN_free(a);
         BN_free(b);
         BN_free(c);
@@ -587,9 +771,7 @@ BN_CTX *ctx;
         return(1);
         }
 
-int test_mod_exp(bp,ctx)
-BIO *bp;
-BN_CTX *ctx;
+int test_mod_exp(BIO *bp, BN_CTX *ctx)
         {
         BIGNUM *a,*b,*c,*d,*e;
         int i;
@@ -600,11 +782,11 @@ BN_CTX *ctx;
         d=BN_new();
         e=BN_new();
 
-        BN_rand(c,30,0,1); /* must be odd for montgomery */
-        for (i=0; i<6; i++)
+        BN_bntest_rand(c,30,0,1); /* must be odd for montgomery */
+        for (i=0; i<num2; i++)
                 {
-                BN_rand(a,20+i*5,0,0); /**/
-                BN_rand(b,2+i,0,0); /**/
+                BN_bntest_rand(a,20+i*5,0,0); /**/
+                BN_bntest_rand(b,2+i,0,0); /**/
 
                 if (!BN_mod_exp(d,a,b,c,ctx))
                         return(00);
@@ -623,6 +805,14 @@ BN_CTX *ctx;
                         BN_print(bp,d);
                         BIO_puts(bp,"\n");
                         }
+                BN_exp(e,a,b,ctx);
+                BN_sub(e,e,d);
+                BN_div(a,b,e,c,ctx);
+                if(!BN_is_zero(b))
+                    {
+                    fprintf(stderr,"Modulo exponentiation test failed!\n");
+                    return 0;
+                    }
                 }
         BN_free(a);
         BN_free(b);
@@ -632,20 +822,252 @@ BN_CTX *ctx;
         return(1);
         }
 
-int test_lshift(bp)
-BIO *bp;
+int test_exp(BIO *bp, BN_CTX *ctx)
         {
-        BIGNUM *a,*b,*c;
+        BIGNUM *a,*b,*d,*e,*one;
         int i;
 
         a=BN_new();
         b=BN_new();
+        d=BN_new();
+        e=BN_new();
+        one=BN_new();
+        BN_one(one);
+
+        for (i=0; i<num2; i++)
+                {
+                BN_bntest_rand(a,20+i*5,0,0); /**/
+                BN_bntest_rand(b,2+i,0,0); /**/
+
+                if (!BN_exp(d,a,b,ctx))
+                        return(00);
+
+                if (bp != NULL)
+                        {
+                        if (!results)
+                                {
+                                BN_print(bp,a);
+                                BIO_puts(bp," ^ ");
+                                BN_print(bp,b);
+                                BIO_puts(bp," - ");
+                                }
+                        BN_print(bp,d);
+                        BIO_puts(bp,"\n");
+                        }
+                BN_one(e);
+                for( ; !BN_is_zero(b) ; BN_sub(b,b,one))
+                    BN_mul(e,e,a,ctx);
+                BN_sub(e,e,d);
+                if(!BN_is_zero(e))
+                    {
+                    fprintf(stderr,"Exponentiation test failed!\n");
+                    return 0;
+                    }
+                }
+        BN_free(a);
+        BN_free(b);
+        BN_free(d);
+        BN_free(e);
+        BN_free(one);
+        return(1);
+        }
+
+static void genprime_cb(int p, int n, void *arg)
+        {
+        char c='*';
+
+        if (p == 0) c='.';
+        if (p == 1) c='+';
+        if (p == 2) c='*';
+        if (p == 3) c='\n';
+        putc(c, stderr);
+        fflush(stderr);
+        (void)n;
+        (void)arg;
+        }
+
+int test_kron(BIO *bp, BN_CTX *ctx)
+        {
+        BIGNUM *a,*b,*r,*t;
+        int i;
+        int legendre, kronecker;
+        int ret = 0;
+
+        a = BN_new();
+        b = BN_new();
+        r = BN_new();
+        t = BN_new();
+        if (a == NULL || b == NULL || r == NULL || t == NULL) goto err;
+        
+        /* We test BN_kronecker(a, b, ctx) just for  b  odd (Jacobi symbol).
+         * In this case we know that if  b  is prime, then BN_kronecker(a, b, ctx)
+         * is congruent to $a^{(b-1)/2}$, modulo $b$ (Legendre symbol).
+         * So we generate a random prime  b  and compare these values
+         * for a number of random  a's.  (That is, we run the Solovay-Strassen
+         * primality test to confirm that  b  is prime, except that we
+         * don't want to test whether  b  is prime but whether BN_kronecker
+         * works.) */
+
+        if (!BN_generate_prime(b, 512, 0, NULL, NULL, genprime_cb, NULL)) goto err;
+        b->neg = rand_neg();
+        putc('\n', stderr);
+
+        for (i = 0; i < num0; i++)
+                {
+                if (!BN_bntest_rand(a, 512, 0, 0)) goto err;
+                a->neg = rand_neg();
+
+                /* t := (|b|-1)/2  (note that b is odd) */
+                if (!BN_copy(t, b)) goto err;
+                t->neg = 0;
+                if (!BN_sub_word(t, 1)) goto err;
+                if (!BN_rshift1(t, t)) goto err;
+                /* r := a^t mod b */
+                b->neg=0;
+                
+                if (!BN_mod_exp_recp(r, a, t, b, ctx)) goto err; /* XXX should be BN_mod_exp_recp, but ..._recp triggers a bug that must be fixed */
+                b->neg=1;
+
+                if (BN_is_word(r, 1))
+                        legendre = 1;
+                else if (BN_is_zero(r))
+                        legendre = 0;
+                else
+                        {
+                        if (!BN_add_word(r, 1)) goto err;
+                        if (0 != BN_ucmp(r, b))
+                                {
+                                fprintf(stderr, "Legendre symbol computation failed\n");
+                                goto err;
+                                }
+                        legendre = -1;
+                        }
+                
+                kronecker = BN_kronecker(a, b, ctx);
+                if (kronecker < -1) goto err;
+                /* we actually need BN_kronecker(a, |b|) */
+                if (a->neg && b->neg)
+                        kronecker = -kronecker;
+                
+                if (legendre != kronecker)
+                        {
+                        fprintf(stderr, "legendre != kronecker; a = ");
+                        BN_print_fp(stderr, a);
+                        fprintf(stderr, ", b = ");
+                        BN_print_fp(stderr, b);
+                        fprintf(stderr, "\n");
+                        goto err;
+                        }
+
+                putc('.', stderr);
+                fflush(stderr);
+                }
+
+        putc('\n', stderr);
+        fflush(stderr);
+        ret = 1;
+ err:
+        if (a != NULL) BN_free(a);
+        if (b != NULL) BN_free(b);
+        if (r != NULL) BN_free(r);
+        if (t != NULL) BN_free(t);
+        return ret;
+        }
+
+int test_sqrt(BIO *bp, BN_CTX *ctx)
+        {
+        BIGNUM *a,*p,*r;
+        int i, j;
+        int ret = 0;
+
+        a = BN_new();
+        p = BN_new();
+        r = BN_new();
+        if (a == NULL || p == NULL || r == NULL) goto err;
+        
+        for (i = 0; i < 16; i++)
+                {
+                if (i < 8)
+                        {
+                        unsigned primes[8] = { 2, 3, 5, 7, 11, 13, 17, 19 };
+                        
+                        if (!BN_set_word(p, primes[i])) goto err;
+                        }
+                else
+                        {
+                        if (!BN_set_word(a, 32)) goto err;
+                        if (!BN_set_word(r, 2*i + 1)) goto err;
+                
+                        if (!BN_generate_prime(p, 256, 0, a, r, genprime_cb, NULL)) goto err;
+                        putc('\n', stderr);
+                        }
+                p->neg = rand_neg();
+
+                for (j = 0; j < num2; j++)
+                        {
+                        /* construct 'a' such that it is a square modulo p,
+                         * but in general not a proper square and not reduced modulo p */
+                        if (!BN_bntest_rand(r, 256, 0, 3)) goto err;
+                        if (!BN_nnmod(r, r, p, ctx)) goto err;
+                        if (!BN_mod_sqr(r, r, p, ctx)) goto err;
+                        if (!BN_bntest_rand(a, 256, 0, 3)) goto err;
+                        if (!BN_nnmod(a, a, p, ctx)) goto err;
+                        if (!BN_mod_sqr(a, a, p, ctx)) goto err;
+                        if (!BN_mul(a, a, r, ctx)) goto err;
+                        if (rand_neg())
+                                if (!BN_sub(a, a, p)) goto err;
+
+                        if (!BN_mod_sqrt(r, a, p, ctx)) goto err;
+                        if (!BN_mod_sqr(r, r, p, ctx)) goto err;
+
+                        if (!BN_nnmod(a, a, p, ctx)) goto err;
+
+                        if (BN_cmp(a, r) != 0)
+                                {
+                                fprintf(stderr, "BN_mod_sqrt failed: a = ");
+                                BN_print_fp(stderr, a);
+                                fprintf(stderr, ", r = ");
+                                BN_print_fp(stderr, r);
+                                fprintf(stderr, ", p = ");
+                                BN_print_fp(stderr, p);
+                                fprintf(stderr, "\n");
+                                goto err;
+                                }
+
+                        putc('.', stderr);
+                        fflush(stderr);
+                        }
+                
+                putc('\n', stderr);
+                fflush(stderr);
+                }
+        ret = 1;
+ err:
+        if (a != NULL) BN_free(a);
+        if (p != NULL) BN_free(p);
+        if (r != NULL) BN_free(r);
+        return ret;
+        }
+
+int test_lshift(BIO *bp,BN_CTX *ctx,BIGNUM *a_)
+        {
+        BIGNUM *a,*b,*c,*d;
+        int i;
+
+        b=BN_new();
         c=BN_new();
+        d=BN_new();
         BN_one(c);
 
-        BN_rand(a,200,0,0); /**/
-        a->neg=rand_neg();
-        for (i=0; i<70; i++)
+        if(a_)
+            a=a_;
+        else
+            {
+            a=BN_new();
+            BN_bntest_rand(a,200,0,0); /**/
+            a->neg=rand_neg();
+            }
+        for (i=0; i<num0; i++)
                 {
                 BN_lshift(b,a,i+1);
                 BN_add(c,c,c);
@@ -661,25 +1083,42 @@ BIO *bp;
                         BN_print(bp,b);
                         BIO_puts(bp,"\n");
                         }
+                BN_mul(d,a,c,ctx);
+                BN_sub(d,d,b);
+                if(!BN_is_zero(d))
+                    {
+                    fprintf(stderr,"Left shift test failed!\n");
+                    fprintf(stderr,"a=");
+                    BN_print_fp(stderr,a);
+                    fprintf(stderr,"\nb=");
+                    BN_print_fp(stderr,b);
+                    fprintf(stderr,"\nc=");
+                    BN_print_fp(stderr,c);
+                    fprintf(stderr,"\nd=");
+                    BN_print_fp(stderr,d);
+                    fprintf(stderr,"\n");
+                    return 0;
+                    }
                 }
         BN_free(a);
         BN_free(b);
         BN_free(c);
+        BN_free(d);
         return(1);
         }
 
-int test_lshift1(bp)
-BIO *bp;
+int test_lshift1(BIO *bp)
         {
-        BIGNUM *a,*b;
+        BIGNUM *a,*b,*c;
         int i;
 
         a=BN_new();
         b=BN_new();
+        c=BN_new();
 
-        BN_rand(a,200,0,0); /**/
+        BN_bntest_rand(a,200,0,0); /**/
         a->neg=rand_neg();
-        for (i=0; i<70; i++)
+        for (i=0; i<num0; i++)
                 {
                 BN_lshift1(b,a);
                 if (bp != NULL)
@@ -693,27 +1132,37 @@ BIO *bp;
                         BN_print(bp,b);
                         BIO_puts(bp,"\n");
                         }
+                BN_add(c,a,a);
+                BN_sub(a,b,c);
+                if(!BN_is_zero(a))
+                    {
+                    fprintf(stderr,"Left shift one test failed!\n");
+                    return 0;
+                    }
+                
                 BN_copy(a,b);
                 }
         BN_free(a);
         BN_free(b);
+        BN_free(c);
         return(1);
         }
 
-int test_rshift(bp)
-BIO *bp;
+int test_rshift(BIO *bp,BN_CTX *ctx)
         {
-        BIGNUM *a,*b,*c;
+        BIGNUM *a,*b,*c,*d,*e;
         int i;
 
         a=BN_new();
         b=BN_new();
         c=BN_new();
+        d=BN_new();
+        e=BN_new();
         BN_one(c);
 
-        BN_rand(a,200,0,0); /**/
+        BN_bntest_rand(a,200,0,0); /**/
         a->neg=rand_neg();
-        for (i=0; i<70; i++)
+        for (i=0; i<num0; i++)
                 {
                 BN_rshift(b,a,i+1);
                 BN_add(c,c,c);
@@ -729,25 +1178,34 @@ BIO *bp;
                         BN_print(bp,b);
                         BIO_puts(bp,"\n");
                         }
+                BN_div(d,e,a,c,ctx);
+                BN_sub(d,d,b);
+                if(!BN_is_zero(d))
+                    {
+                    fprintf(stderr,"Right shift test failed!\n");
+                    return 0;
+                    }
                 }
         BN_free(a);
         BN_free(b);
         BN_free(c);
+        BN_free(d);
+        BN_free(e);
         return(1);
         }
 
-int test_rshift1(bp)
-BIO *bp;
+int test_rshift1(BIO *bp)
         {
-        BIGNUM *a,*b;
+        BIGNUM *a,*b,*c;
         int i;
 
         a=BN_new();
         b=BN_new();
+        c=BN_new();
 
-        BN_rand(a,200,0,0); /**/
+        BN_bntest_rand(a,200,0,0); /**/
         a->neg=rand_neg();
-        for (i=0; i<70; i++)
+        for (i=0; i<num0; i++)
                 {
                 BN_rshift1(b,a);
                 if (bp != NULL)
@@ -761,14 +1219,22 @@ BIO *bp;
                         BN_print(bp,b);
                         BIO_puts(bp,"\n");
                         }
+                BN_sub(c,a,b);
+                BN_sub(c,c,b);
+                if(!BN_is_zero(c) && !BN_abs_is_word(c, 1))
+                    {
+                    fprintf(stderr,"Right shift one test failed!\n");
+                    return 0;
+                    }
                 BN_copy(a,b);
                 }
         BN_free(a);
         BN_free(b);
+        BN_free(c);
         return(1);
         }
 
-int rand_neg()
+int rand_neg(void)
         {
         static unsigned int neg=0;
         static int sign[8]={0,0,0,1,1,0,1,1};
diff --git a/src/lib/libcrypto/bn/exp.c b/src/lib/libcrypto/bn/exp.c
index ec443459d8..4865b0ef74 100644
--- a/src/lib/libcrypto/bn/exp.c
+++ b/src/lib/libcrypto/bn/exp.c
@@ -1,3 +1,5 @@
+/* unused */
+
 #include <stdio.h>
 #include <openssl/tmdiff.h>
 #include "bn_lcl.h"
diff --git a/src/lib/libcrypto/bn/expspeed.c b/src/lib/libcrypto/bn/expspeed.c
index 344f883d35..07a1bcf51c 100644
--- a/src/lib/libcrypto/bn/expspeed.c
+++ b/src/lib/libcrypto/bn/expspeed.c
@@ -1,3 +1,5 @@
+/* unused */
+
 /* crypto/bn/expspeed.c */
 /* Copyright (C) 1995-1998 Eric Young (eay@cryptsoft.com)
  * All rights reserved.
@@ -59,6 +61,31 @@
 /* most of this code has been pilfered from my libdes speed.c program */
 
 #define BASENUM 5000
+#define NUM_START 0
+
+
+/* determine timings for modexp, modmul, modsqr, gcd, Kronecker symbol,
+ * modular inverse, or modular square roots */
+#define TEST_EXP
+#undef TEST_MUL
+#undef TEST_SQR
+#undef TEST_GCD
+#undef TEST_KRON
+#undef TEST_INV
+#undef TEST_SQRT
+#define P_MOD_64 9 /* least significant 6 bits for prime to be used for BN_sqrt timings */
+
+#if defined(TEST_EXP) + defined(TEST_MUL) + defined(TEST_SQR) + defined(TEST_GCD) + defined(TEST_KRON) + defined(TEST_INV) +defined(TEST_SQRT) != 1
+#  error "choose one test"
+#endif
+
+#if defined(TEST_INV) || defined(TEST_SQRT)
+#  define C_PRIME
+static void genprime_cb(int p, int n, void *arg);
+#endif
+
+
+
 #undef PROG
 #define PROG bnspeed_main
 
@@ -66,14 +93,14 @@
 #include <stdlib.h>
 #include <signal.h>
 #include <string.h>
-#include "crypto.h"
-#include "err.h"
+#include <openssl/crypto.h>
+#include <openssl/err.h>
+#include <openssl/rand.h>
 
-#ifndef MSDOS
+#if !defined(OPENSSL_SYS_MSDOS) && (!defined(OPENSSL_SYS_VMS) || defined(__DECC)) && !defined(OPENSSL_SYS_MACOSX)
 #define TIMES
 #endif
 
-#ifndef VMS
 #ifndef _IRIX
 #include <time.h>
 #endif
@@ -81,36 +108,33 @@
 #include <sys/types.h>
 #include <sys/times.h>
 #endif
-#else /* VMS */
-#include <types.h>
-struct tms {
-        time_t tms_utime;
-        time_t tms_stime;
-        time_t tms_uchild;      /* I dunno...  */
-        time_t tms_uchildsys;   /* so these names are a guess :-) */
-        }
+
+/* Depending on the VMS version, the tms structure is perhaps defined.
+   The __TMS macro will show if it was.  If it wasn't defined, we should
+   undefine TIMES, since that tells the rest of the program how things
+   should be handled.                           -- Richard Levitte */
+#if defined(OPENSSL_SYS_VMS_DECC) && !defined(__TMS)
+#undef TIMES
 #endif
+
 #ifndef TIMES
 #include <sys/timeb.h>
 #endif
 
-#ifdef sun
+#if defined(sun) || defined(__ultrix)
+#define _POSIX_SOURCE
 #include <limits.h>
 #include <sys/param.h>
 #endif
 
-#include "bn.h"
-#include "x509.h"
+#include <openssl/bn.h>
+#include <openssl/x509.h>
 
 /* The following if from times(3) man page.  It may need to be changed */
 #ifndef HZ
 # ifndef CLK_TCK
 #  ifndef _BSD_CLK_TCK_ /* FreeBSD hack */
-#   ifndef VMS
-#    define HZ  100.0
-#   else /* VMS */
-#    define HZ  100.0
-#   endif
+#   define HZ   100.0
 #  else /* _BSD_CLK_TCK_ */
 #   define HZ ((double)_BSD_CLK_TCK_)
 #  endif
@@ -123,17 +147,11 @@ struct tms {
 #define BUFSIZE ((long)1024*8)
 int run=0;
 
-#ifndef NOPROTO
 static double Time_F(int s);
-#else
-static double Time_F();
-#endif
-
 #define START   0
 #define STOP    1
 
-static double Time_F(s)
-int s;
+static double Time_F(int s)
         {
         double ret;
 #ifdef TIMES
@@ -169,62 +187,167 @@ int s;
 #endif
         }
 
-#define NUM_SIZES       6
-static int sizes[NUM_SIZES]={256,512,1024,2048,4096,8192};
-static int mul_c[NUM_SIZES]={8*8*8*8*8,8*8*8*8,8*8*8,8*8,8,1};
+#define NUM_SIZES       7
+#if NUM_START > NUM_SIZES
+#   error "NUM_START > NUM_SIZES"
+#endif
+static int sizes[NUM_SIZES]={128,256,512,1024,2048,4096,8192};
+static int mul_c[NUM_SIZES]={8*8*8*8*8*8,8*8*8*8*8,8*8*8*8,8*8*8,8*8,8,1};
 /*static int sizes[NUM_SIZES]={59,179,299,419,539}; */
 
+#define RAND_SEED(string) { const char str[] = string; RAND_seed(string, sizeof str); }
+
 void do_mul_exp(BIGNUM *r,BIGNUM *a,BIGNUM *b,BIGNUM *c,BN_CTX *ctx); 
 
-int main(argc,argv)
-int argc;
-char **argv;
+int main(int argc, char **argv)
         {
         BN_CTX *ctx;
         BIGNUM *a,*b,*c,*r;
 
+#if 1
+        if (!CRYPTO_set_mem_debug_functions(0,0,0,0,0))
+                abort();
+#endif
+
         ctx=BN_CTX_new();
         a=BN_new();
         b=BN_new();
         c=BN_new();
         r=BN_new();
 
+        while (!RAND_status())
+                /* not enough bits */
+                RAND_SEED("I demand a manual recount!");
+
         do_mul_exp(r,a,b,c,ctx);
+        return 0;
         }
 
-void do_mul_exp(r,a,b,c,ctx)
-BIGNUM *r;
-BIGNUM *a;
-BIGNUM *b;
-BIGNUM *c;
-BN_CTX *ctx;
+void do_mul_exp(BIGNUM *r, BIGNUM *a, BIGNUM *b, BIGNUM *c, BN_CTX *ctx)
         {
         int i,k;
         double tm;
         long num;
-        BN_MONT_CTX m;
-
-        memset(&m,0,sizeof(m));
 
         num=BASENUM;
-        for (i=0; i<NUM_SIZES; i++)
+        for (i=NUM_START; i<NUM_SIZES; i++)
                 {
-                BN_rand(a,sizes[i],1,0);
-                BN_rand(b,sizes[i],1,0);
-                BN_rand(c,sizes[i],1,1);
-                BN_mod(a,a,c,ctx);
-                BN_mod(b,b,c,ctx);
-
-                BN_MONT_CTX_set(&m,c,ctx);
+#ifdef C_PRIME
+#  ifdef TEST_SQRT
+                if (!BN_set_word(a, 64)) goto err;
+                if (!BN_set_word(b, P_MOD_64)) goto err;
+#    define ADD a
+#    define REM b
+#  else
+#    define ADD NULL
+#    define REM NULL
+#  endif
+                if (!BN_generate_prime(c,sizes[i],0,ADD,REM,genprime_cb,NULL)) goto err;
+                putc('\n', stderr);
+                fflush(stderr);
+#endif
 
-                Time_F(START);
                 for (k=0; k<num; k++)
-                        BN_mod_exp_mont(r,a,b,c,ctx,&m);
+                        {
+                        if (k%50 == 0) /* Average over num/50 different choices of random numbers. */
+                                {
+                                if (!BN_pseudo_rand(a,sizes[i],1,0)) goto err;
+
+                                if (!BN_pseudo_rand(b,sizes[i],1,0)) goto err;
+
+#ifndef C_PRIME
+                                if (!BN_pseudo_rand(c,sizes[i],1,1)) goto err;
+#endif
+
+#ifdef TEST_SQRT                                
+                                if (!BN_mod_sqr(a,a,c,ctx)) goto err;
+                                if (!BN_mod_sqr(b,b,c,ctx)) goto err;
+#else
+                                if (!BN_nnmod(a,a,c,ctx)) goto err;
+                                if (!BN_nnmod(b,b,c,ctx)) goto err;
+#endif
+
+                                if (k == 0)
+                                        Time_F(START);
+                                }
+
+#if defined(TEST_EXP)
+                        if (!BN_mod_exp(r,a,b,c,ctx)) goto err;
+#elif defined(TEST_MUL)
+                        {
+                        int i = 0;
+                        for (i = 0; i < 50; i++)
+                                if (!BN_mod_mul(r,a,b,c,ctx)) goto err;
+                        }
+#elif defined(TEST_SQR)
+                        {
+                        int i = 0;
+                        for (i = 0; i < 50; i++)
+                                {
+                                if (!BN_mod_sqr(r,a,c,ctx)) goto err;
+                                if (!BN_mod_sqr(r,b,c,ctx)) goto err;
+                                }
+                        }
+#elif defined(TEST_GCD)
+                        if (!BN_gcd(r,a,b,ctx)) goto err;
+                        if (!BN_gcd(r,b,c,ctx)) goto err;
+                        if (!BN_gcd(r,c,a,ctx)) goto err;
+#elif defined(TEST_KRON)
+                        if (-2 == BN_kronecker(a,b,ctx)) goto err;
+                        if (-2 == BN_kronecker(b,c,ctx)) goto err;
+                        if (-2 == BN_kronecker(c,a,ctx)) goto err;
+#elif defined(TEST_INV)
+                        if (!BN_mod_inverse(r,a,c,ctx)) goto err;
+                        if (!BN_mod_inverse(r,b,c,ctx)) goto err;
+#else /* TEST_SQRT */
+                        if (!BN_mod_sqrt(r,a,c,ctx)) goto err;
+                        if (!BN_mod_sqrt(r,b,c,ctx)) goto err;
+#endif
+                        }
                 tm=Time_F(STOP);
-                printf("mul %4d ^ %4d %% %d -> %8.3fms %5.1f\n",sizes[i],sizes[i],sizes[i],tm*1000.0/num,tm*mul_c[i]/num);
+                printf(
+#if defined(TEST_EXP)
+                        "modexp %4d ^ %4d %% %4d"
+#elif defined(TEST_MUL)
+                        "50*modmul %4d %4d %4d"
+#elif defined(TEST_SQR)
+                        "100*modsqr %4d %4d %4d"
+#elif defined(TEST_GCD)
+                        "3*gcd %4d %4d %4d"
+#elif defined(TEST_KRON)
+                        "3*kronecker %4d %4d %4d"
+#elif defined(TEST_INV)
+                        "2*inv %4d %4d mod %4d"
+#else /* TEST_SQRT */
+                        "2*sqrt [prime == %d (mod 64)] %4d %4d mod %4d"
+#endif
+                        " -> %8.3fms %5.1f (%ld)\n",
+#ifdef TEST_SQRT
+                        P_MOD_64,
+#endif
+                        sizes[i],sizes[i],sizes[i],tm*1000.0/num,tm*mul_c[i]/num, num);
                 num/=7;
                 if (num <= 0) num=1;
                 }
+        return;
 
+ err:
+        ERR_print_errors_fp(stderr);
         }
 
+
+#ifdef C_PRIME
+static void genprime_cb(int p, int n, void *arg)
+        {
+        char c='*';
+
+        if (p == 0) c='.';
+        if (p == 1) c='+';
+        if (p == 2) c='*';
+        if (p == 3) c='\n';
+        putc(c, stderr);
+        fflush(stderr);
+        (void)n;
+        (void)arg;
+        }
+#endif
diff --git a/src/lib/libcrypto/bn/exptest.c b/src/lib/libcrypto/bn/exptest.c
index 67dc95d726..5ca570d1a8 100644
--- a/src/lib/libcrypto/bn/exptest.c
+++ b/src/lib/libcrypto/bn/exptest.c
@@ -59,30 +59,37 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
-#include "bio.h"
-#include "bn.h"
-#include "rand.h"
-#include "err.h"
-#ifdef WINDOWS
+#include <openssl/bio.h>
+#include <openssl/bn.h>
+#include <openssl/rand.h>
+#include <openssl/err.h>
+#ifdef OPENSSL_SYS_WINDOWS
 #include "../bio/bss_file.c"
 #endif
 
 #define NUM_BITS        (BN_BITS*2)
 
-int main(argc,argv)
-int argc;
-char *argv[];
+static const char rnd_seed[] = "string to make the random number generator think it has entropy";
+
+int main(int argc, char *argv[])
         {
         BN_CTX *ctx;
         BIO *out=NULL;
         int i,ret;
         unsigned char c;
-        BIGNUM *r_mont,*r_recp,*a,*b,*m;
+        BIGNUM *r_mont,*r_recp,*r_simple,*a,*b,*m;
+
+        RAND_seed(rnd_seed, sizeof rnd_seed); /* or BN_rand may fail, and we don't
+                                               * even check its return value
+                                               * (which we should) */
+
+        ERR_load_BN_strings();
 
         ctx=BN_CTX_new();
         if (ctx == NULL) exit(1);
         r_mont=BN_new();
         r_recp=BN_new();
+        r_simple=BN_new();
         a=BN_new();
         b=BN_new();
         m=BN_new();
@@ -114,29 +121,61 @@ char *argv[];
 
                 ret=BN_mod_exp_mont(r_mont,a,b,m,ctx,NULL);
                 if (ret <= 0)
-                        { printf("BN_mod_exp_mont() problems\n"); exit(1); }
+                        {
+                        printf("BN_mod_exp_mont() problems\n");
+                        ERR_print_errors(out);
+                        exit(1);
+                        }
 
                 ret=BN_mod_exp_recp(r_recp,a,b,m,ctx);
                 if (ret <= 0)
-                        { printf("BN_mod_exp_recp() problems\n"); exit(1); }
-                
-                if (BN_cmp(r_mont,r_recp) != 0)
                         {
-                        printf("\nmont and recp results differ\n");
+                        printf("BN_mod_exp_recp() problems\n");
+                        ERR_print_errors(out);
+                        exit(1);
+                        }
+
+                ret=BN_mod_exp_simple(r_simple,a,b,m,ctx);
+                if (ret <= 0)
+                        {
+                        printf("BN_mod_exp_simple() problems\n");
+                        ERR_print_errors(out);
+                        exit(1);
+                        }
+
+                if (BN_cmp(r_simple, r_mont) == 0
+                    && BN_cmp(r_simple,r_recp) == 0)
+                        {
+                        printf(".");
+                        fflush(stdout);
+                        }
+                else
+                        {
+                        if (BN_cmp(r_simple,r_mont) != 0)
+                                printf("\nsimple and mont results differ\n");
+                        if (BN_cmp(r_simple,r_recp) != 0)
+                                printf("\nsimple and recp results differ\n");
+
                         printf("a (%3d) = ",BN_num_bits(a));   BN_print(out,a);
                         printf("\nb (%3d) = ",BN_num_bits(b)); BN_print(out,b);
                         printf("\nm (%3d) = ",BN_num_bits(m)); BN_print(out,m);
+                        printf("\nsimple   ="); BN_print(out,r_simple);
                         printf("\nrecp     ="); BN_print(out,r_recp);
                         printf("\nmont     ="); BN_print(out,r_mont);
                         printf("\n");
                         exit(1);
                         }
-                else
-                        {
-                        printf(".");
-                        fflush(stdout);
-                        }
                 }
+        BN_free(r_mont);
+        BN_free(r_recp);
+        BN_free(r_simple);
+        BN_free(a);
+        BN_free(b);
+        BN_free(m);
+        BN_CTX_free(ctx);
+        ERR_remove_state(0);
+        CRYPTO_mem_leaks(out);
+        BIO_free(out);
         printf(" done\n");
         exit(0);
 err:
diff --git a/src/lib/libcrypto/bn/vms-helper.c b/src/lib/libcrypto/bn/vms-helper.c
index 73af337069..4b63149bf3 100644
--- a/src/lib/libcrypto/bn/vms-helper.c
+++ b/src/lib/libcrypto/bn/vms-helper.c
@@ -59,8 +59,10 @@
 
 bn_div_words_abort(int i)
 {
-#if !defined(NO_STDIO) && !defined(WIN16)
+#ifdef BN_DEBUG
+#if !defined(OPENSSL_NO_STDIO) && !defined(OPENSSL_SYS_WIN16)
         fprintf(stderr,"Division would overflow (%d)\n",i);
 #endif
         abort();
+#endif
 }