OpenSSL 0.9.4 merge

18 files changed, 17172 insertions, 1399 deletions

diff --git a/src/lib/libcrypto/bn/asm/alpha.s b/src/lib/libcrypto/bn/asm/alpha.s
index 1d17b1d619..a351694ca2 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
+        bgt     $19,$945
+        ret     $31,($26),1     # else exit
 
-        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--
-
-        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 +453,1446 @@ $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
+
+        subq    $30,    16,     $30
+        ldq     $0,     0($17)
+        ldq     $1,     0($18)
+        stq     $9,     0($30)
+        stq     $10,    8($30)
+        ldq     $2,     8($17)
+        ldq     $3,     8($18)
+        ldq     $4,     16($17)
+        ldq     $5,     16($18)
+        ldq     $6,     24($17)
+        ldq     $7,     24($18)
+        ldq     $8,     8($17)
+        ldq     $22,    8($18)
+        ldq     $23,    8($17)
+        ldq     $24,    8($18)
+        ldq     $25,    8($17)
+        ldq     $27,    8($18)
+        ldq     $28,    8($17)
+        ldq     $21,    8($18)
+        bis     $31,    $31,    $9
+        mulq    $0,     $1,     $20
+        umulh   $0,     $1,     $19
+        stq     $20,    0($16)
+        bis     $31,    $31,    $20
+        mulq    $0,     $3,     $10
+        umulh   $0,     $3,     $17
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $10
+        addq    $20,    $10,    $20
+        mulq    $2,     $1,     $18
+        umulh   $2,     $1,     $17
+        addq    $19,    $18,    $19
+        cmpult  $19,    $18,    $10
+        addq    $10,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $18
+        addq    $20,    $18,    $20
+        stq     $19,    8($16)
+        bis     $31,    $31,    $19
+        mulq    $0,     $5,     $10
+        umulh   $0,     $5,     $17
+        addq    $9,     $10,    $9
+        cmpult  $9,     $10,    $18
+        addq    $18,    $17,    $17
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $10
+        addq    $19,    $10,    $19
+        mulq    $2,     $3,     $18
+        umulh   $2,     $3,     $17
+        addq    $9,     $18,    $9
+        cmpult  $9,     $18,    $10
+        addq    $10,    $17,    $17
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $18
+        addq    $19,    $18,    $19
+        mulq    $4,     $1,     $10
+        umulh   $4,     $1,     $17
+        addq    $9,     $10,    $9
+        cmpult  $9,     $10,    $18
+        addq    $18,    $17,    $17
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $10
+        addq    $19,    $10,    $19
+        stq     $9,     16($16)
+        bis     $31,    $31,    $9
+        mulq    $0,     $7,     $18
+        umulh   $0,     $7,     $17
+        addq    $20,    $18,    $20
+        cmpult  $20,    $18,    $10
+        addq    $10,    $17,    $17
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $18
+        addq    $9,     $18,    $9
+        mulq    $2,     $5,     $10
+        umulh   $2,     $5,     $17
+        addq    $20,    $10,    $20
+        cmpult  $20,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $10
+        addq    $9,     $10,    $9
+        mulq    $4,     $3,     $18
+        umulh   $4,     $3,     $17
+        addq    $20,    $18,    $20
+        cmpult  $20,    $18,    $10
+        addq    $10,    $17,    $17
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $18
+        addq    $9,     $18,    $9
+        mulq    $6,     $1,     $10
+        umulh   $6,     $1,     $17
+        addq    $20,    $10,    $20
+        cmpult  $20,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $10
+        addq    $9,     $10,    $9
+        stq     $20,    24($16)
+        bis     $31,    $31,    $20
+        mulq    $0,     $22,    $18
+        umulh   $0,     $22,    $17
+        addq    $19,    $18,    $19
+        cmpult  $19,    $18,    $10
+        addq    $10,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $18
+        addq    $20,    $18,    $20
+        mulq    $2,     $7,     $10
+        umulh   $2,     $7,     $17
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $10
+        addq    $20,    $10,    $20
+        mulq    $4,     $5,     $18
+        umulh   $4,     $5,     $17
+        addq    $19,    $18,    $19
+        cmpult  $19,    $18,    $10
+        addq    $10,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $18
+        addq    $20,    $18,    $20
+        mulq    $6,     $3,     $10
+        umulh   $6,     $3,     $17
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $10
+        addq    $20,    $10,    $20
+        mulq    $8,     $1,     $18
+        umulh   $8,     $1,     $17
+        addq    $19,    $18,    $19
+        cmpult  $19,    $18,    $10
+        addq    $10,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $18
+        addq    $20,    $18,    $20
+        stq     $19,    32($16)
+        bis     $31,    $31,    $19
+        mulq    $0,     $24,    $10
+        umulh   $0,     $24,    $17
+        addq    $9,     $10,    $9
+        cmpult  $9,     $10,    $18
+        addq    $18,    $17,    $17
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $10
+        addq    $19,    $10,    $19
+        mulq    $2,     $22,    $18
+        umulh   $2,     $22,    $17
+        addq    $9,     $18,    $9
+        cmpult  $9,     $18,    $10
+        addq    $10,    $17,    $17
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $18
+        addq    $19,    $18,    $19
+        mulq    $4,     $7,     $10
+        umulh   $4,     $7,     $17
+        addq    $9,     $10,    $9
+        cmpult  $9,     $10,    $18
+        addq    $18,    $17,    $17
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $10
+        addq    $19,    $10,    $19
+        mulq    $6,     $5,     $18
+        umulh   $6,     $5,     $17
+        addq    $9,     $18,    $9
+        cmpult  $9,     $18,    $10
+        addq    $10,    $17,    $17
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $18
+        addq    $19,    $18,    $19
+        mulq    $8,     $3,     $10
+        umulh   $8,     $3,     $17
+        addq    $9,     $10,    $9
+        cmpult  $9,     $10,    $18
+        addq    $18,    $17,    $17
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $10
+        addq    $19,    $10,    $19
+        mulq    $23,    $1,     $18
+        umulh   $23,    $1,     $17
+        addq    $9,     $18,    $9
+        cmpult  $9,     $18,    $10
+        addq    $10,    $17,    $17
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $18
+        addq    $19,    $18,    $19
+        stq     $9,     40($16)
+        bis     $31,    $31,    $9
+        mulq    $0,     $27,    $10
+        umulh   $0,     $27,    $17
+        addq    $20,    $10,    $20
+        cmpult  $20,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $10
+        addq    $9,     $10,    $9
+        mulq    $2,     $24,    $18
+        umulh   $2,     $24,    $17
+        addq    $20,    $18,    $20
+        cmpult  $20,    $18,    $10
+        addq    $10,    $17,    $17
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $18
+        addq    $9,     $18,    $9
+        mulq    $4,     $22,    $10
+        umulh   $4,     $22,    $17
+        addq    $20,    $10,    $20
+        cmpult  $20,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $10
+        addq    $9,     $10,    $9
+        mulq    $6,     $7,     $18
+        umulh   $6,     $7,     $17
+        addq    $20,    $18,    $20
+        cmpult  $20,    $18,    $10
+        addq    $10,    $17,    $17
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $18
+        addq    $9,     $18,    $9
+        mulq    $8,     $5,     $10
+        umulh   $8,     $5,     $17
+        addq    $20,    $10,    $20
+        cmpult  $20,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $10
+        addq    $9,     $10,    $9
+        mulq    $23,    $3,     $18
+        umulh   $23,    $3,     $17
+        addq    $20,    $18,    $20
+        cmpult  $20,    $18,    $10
+        addq    $10,    $17,    $17
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $18
+        addq    $9,     $18,    $9
+        mulq    $25,    $1,     $10
+        umulh   $25,    $1,     $17
+        addq    $20,    $10,    $20
+        cmpult  $20,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $10
+        addq    $9,     $10,    $9
+        stq     $20,    48($16)
+        bis     $31,    $31,    $20
+        mulq    $0,     $21,    $18
+        umulh   $0,     $21,    $17
+        addq    $19,    $18,    $19
+        cmpult  $19,    $18,    $10
+        addq    $10,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $18
+        addq    $20,    $18,    $20
+        mulq    $2,     $27,    $10
+        umulh   $2,     $27,    $17
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $0
+        addq    $20,    $0,     $20
+        mulq    $4,     $24,    $10
+        umulh   $4,     $24,    $18
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $17
+        addq    $17,    $18,    $18
+        addq    $9,     $18,    $9
+        cmpult  $9,     $18,    $0
+        addq    $20,    $0,     $20
+        mulq    $6,     $22,    $10
+        umulh   $6,     $22,    $17
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $0
+        addq    $20,    $0,     $20
+        mulq    $8,     $7,     $10
+        umulh   $8,     $7,     $18
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $17
+        addq    $17,    $18,    $18
+        addq    $9,     $18,    $9
+        cmpult  $9,     $18,    $0
+        addq    $20,    $0,     $20
+        mulq    $23,    $5,     $10
+        umulh   $23,    $5,     $17
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $0
+        addq    $20,    $0,     $20
+        mulq    $25,    $3,     $10
+        umulh   $25,    $3,     $18
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $17
+        addq    $17,    $18,    $18
+        addq    $9,     $18,    $9
+        cmpult  $9,     $18,    $0
+        addq    $20,    $0,     $20
+        mulq    $28,    $1,     $10
+        umulh   $28,    $1,     $17
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $18
+        addq    $18,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $0
+        addq    $20,    $0,     $20
+        stq     $19,    56($16)
+        bis     $31,    $31,    $19
+        mulq    $2,     $21,    $10
+        umulh   $2,     $21,    $18
+        addq    $9,     $10,    $9
+        cmpult  $9,     $10,    $17
+        addq    $17,    $18,    $18
+        addq    $20,    $18,    $20
+        cmpult  $20,    $18,    $0
+        addq    $19,    $0,     $19
+        mulq    $4,     $27,    $1
+        umulh   $4,     $27,    $10
+        addq    $9,     $1,     $9
+        cmpult  $9,     $1,     $17
+        addq    $17,    $10,    $10
+        addq    $20,    $10,    $20
+        cmpult  $20,    $10,    $18
+        addq    $19,    $18,    $19
+        mulq    $6,     $24,    $0
+        umulh   $6,     $24,    $2
+        addq    $9,     $0,     $9
+        cmpult  $9,     $0,     $1
+        addq    $1,     $2,     $2
+        addq    $20,    $2,     $20
+        cmpult  $20,    $2,     $17
+        addq    $19,    $17,    $19
+        mulq    $8,     $22,    $10
+        umulh   $8,     $22,    $18
+        addq    $9,     $10,    $9
+        cmpult  $9,     $10,    $0
+        addq    $0,     $18,    $18
+        addq    $20,    $18,    $20
+        cmpult  $20,    $18,    $1
+        addq    $19,    $1,     $19
+        mulq    $23,    $7,     $2
+        umulh   $23,    $7,     $17
+        addq    $9,     $2,     $9
+        cmpult  $9,     $2,     $10
+        addq    $10,    $17,    $17
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $0
+        addq    $19,    $0,     $19
+        mulq    $25,    $5,     $18
+        umulh   $25,    $5,     $1
+        addq    $9,     $18,    $9
+        cmpult  $9,     $18,    $2
+        addq    $2,     $1,     $1
+        addq    $20,    $1,     $20
+        cmpult  $20,    $1,     $10
+        addq    $19,    $10,    $19
+        mulq    $28,    $3,     $17
+        umulh   $28,    $3,     $0
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $18
+        addq    $18,    $0,     $0
+        addq    $20,    $0,     $20
+        cmpult  $20,    $0,     $2
+        addq    $19,    $2,     $19
+        stq     $9,     64($16)
+        bis     $31,    $31,    $9
+        mulq    $4,     $21,    $1
+        umulh   $4,     $21,    $10
+        addq    $20,    $1,     $20
+        cmpult  $20,    $1,     $17
+        addq    $17,    $10,    $10
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $18
+        addq    $9,     $18,    $9
+        mulq    $6,     $27,    $0
+        umulh   $6,     $27,    $2
+        addq    $20,    $0,     $20
+        cmpult  $20,    $0,     $3
+        addq    $3,     $2,     $2
+        addq    $19,    $2,     $19
+        cmpult  $19,    $2,     $1
+        addq    $9,     $1,     $9
+        mulq    $8,     $24,    $17
+        umulh   $8,     $24,    $10
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $18
+        addq    $18,    $10,    $10
+        addq    $19,    $10,    $19
+        cmpult  $19,    $10,    $4
+        addq    $9,     $4,     $9
+        mulq    $23,    $22,    $0
+        umulh   $23,    $22,    $3
+        addq    $20,    $0,     $20
+        cmpult  $20,    $0,     $2
+        addq    $2,     $3,     $3
+        addq    $19,    $3,     $19
+        cmpult  $19,    $3,     $1
+        addq    $9,     $1,     $9
+        mulq    $25,    $7,     $17
+        umulh   $25,    $7,     $18
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $10
+        addq    $10,    $18,    $18
+        addq    $19,    $18,    $19
+        cmpult  $19,    $18,    $4
+        addq    $9,     $4,     $9
+        mulq    $28,    $5,     $0
+        umulh   $28,    $5,     $2
+        addq    $20,    $0,     $20
+        cmpult  $20,    $0,     $3
+        addq    $3,     $2,     $2
+        addq    $19,    $2,     $19
+        cmpult  $19,    $2,     $1
+        addq    $9,     $1,     $9
+        stq     $20,    72($16)
+        bis     $31,    $31,    $20
+        mulq    $6,     $21,    $17
+        umulh   $6,     $21,    $10
+        addq    $19,    $17,    $19
+        cmpult  $19,    $17,    $18
+        addq    $18,    $10,    $10
+        addq    $9,     $10,    $9
+        cmpult  $9,     $10,    $4
+        addq    $20,    $4,     $20
+        mulq    $8,     $27,    $0
+        umulh   $8,     $27,    $3
+        addq    $19,    $0,     $19
+        cmpult  $19,    $0,     $2
+        addq    $2,     $3,     $3
+        addq    $9,     $3,     $9
+        cmpult  $9,     $3,     $1
+        addq    $20,    $1,     $20
+        mulq    $23,    $24,    $5
+        umulh   $23,    $24,    $17
+        addq    $19,    $5,     $19
+        cmpult  $19,    $5,     $18
+        addq    $18,    $17,    $17
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $10
+        addq    $20,    $10,    $20
+        mulq    $25,    $22,    $4
+        umulh   $25,    $22,    $6
+        addq    $19,    $4,     $19
+        cmpult  $19,    $4,     $0
+        addq    $0,     $6,     $6
+        addq    $9,     $6,     $9
+        cmpult  $9,     $6,     $2
+        addq    $20,    $2,     $20
+        mulq    $28,    $7,     $3
+        umulh   $28,    $7,     $1
+        addq    $19,    $3,     $19
+        cmpult  $19,    $3,     $5
+        addq    $5,     $1,     $1
+        addq    $9,     $1,     $9
+        cmpult  $9,     $1,     $18
+        addq    $20,    $18,    $20
+        stq     $19,    80($16)
+        bis     $31,    $31,    $19
+        mulq    $8,     $21,    $17
+        umulh   $8,     $21,    $10
+        addq    $9,     $17,    $9
+        cmpult  $9,     $17,    $4
+        addq    $4,     $10,    $10
+        addq    $20,    $10,    $20
+        cmpult  $20,    $10,    $0
+        addq    $19,    $0,     $19
+        mulq    $23,    $27,    $6
+        umulh   $23,    $27,    $2
+        addq    $9,     $6,     $9
+        cmpult  $9,     $6,     $3
+        addq    $3,     $2,     $2
+        addq    $20,    $2,     $20
+        cmpult  $20,    $2,     $5
+        addq    $19,    $5,     $19
+        mulq    $25,    $24,    $1
+        umulh   $25,    $24,    $18
+        addq    $9,     $1,     $9
+        cmpult  $9,     $1,     $7
+        addq    $7,     $18,    $18
+        addq    $20,    $18,    $20
+        cmpult  $20,    $18,    $17
+        addq    $19,    $17,    $19
+        mulq    $28,    $22,    $4
+        umulh   $28,    $22,    $10
+        addq    $9,     $4,     $9
+        cmpult  $9,     $4,     $0
+        addq    $0,     $10,    $10
+        addq    $20,    $10,    $20
+        cmpult  $20,    $10,    $8
+        addq    $19,    $8,     $19
+        stq     $9,     88($16)
+        bis     $31,    $31,    $9
+        mulq    $23,    $21,    $6
+        umulh   $23,    $21,    $3
+        addq    $20,    $6,     $20
+        cmpult  $20,    $6,     $2
+        addq    $2,     $3,     $3
+        addq    $19,    $3,     $19
+        cmpult  $19,    $3,     $5
+        addq    $9,     $5,     $9
+        mulq    $25,    $27,    $1
+        umulh   $25,    $27,    $7
+        addq    $20,    $1,     $20
+        cmpult  $20,    $1,     $18
+        addq    $18,    $7,     $7
+        addq    $19,    $7,     $19
+        cmpult  $19,    $7,     $17
+        addq    $9,     $17,    $9
+        mulq    $28,    $24,    $4
+        umulh   $28,    $24,    $0
+        addq    $20,    $4,     $20
+        cmpult  $20,    $4,     $10
+        addq    $10,    $0,     $0
+        addq    $19,    $0,     $19
+        cmpult  $19,    $0,     $8
+        addq    $9,     $8,     $9
+        stq     $20,    96($16)
+        bis     $31,    $31,    $20
+        mulq    $25,    $21,    $22
+        umulh   $25,    $21,    $6
+        addq    $19,    $22,    $19
+        cmpult  $19,    $22,    $2
+        addq    $2,     $6,     $6
+        addq    $9,     $6,     $9
+        cmpult  $9,     $6,     $3
+        addq    $20,    $3,     $20
+        mulq    $28,    $27,    $5
+        umulh   $28,    $27,    $23
+        addq    $19,    $5,     $19
+        cmpult  $19,    $5,     $1
+        addq    $1,     $23,    $23
+        addq    $9,     $23,    $9
+        cmpult  $9,     $23,    $18
+        addq    $20,    $18,    $20
+        stq     $19,    104($16)
+        bis     $31,    $31,    $19
+        mulq    $28,    $21,    $7
+        umulh   $28,    $21,    $17
+        addq    $9,     $7,     $9
+        cmpult  $9,     $7,     $4
+        addq    $4,     $17,    $17
+        addq    $20,    $17,    $20
+        cmpult  $20,    $17,    $10
+        addq    $19,    $10,    $19
+        stq     $9,     112($16)
+        stq     $20,    120($16)
+        ldq     $9,     0($30)
+        ldq     $10,    8($30)
+        addq    $30,    16,     $30
+        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/alpha.s.works b/src/lib/libcrypto/bn/asm/alpha.s.works
new file mode 100644
index 0000000000..ee6c587809
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/alpha.s.works
@@ -0,0 +1,533 @@
+
+ # DEC Alpha assember
+ # The bn_div64 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.
+ # 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:
+        .text
+        .align 3
+        .globl bn_mul_add_words
+        .ent bn_mul_add_words
+bn_mul_add_words:
+bn_mul_add_words..ng:
+        .frame $30,0,$26,0
+        .prologue 0
+        .align 5
+        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
+
+        ldq     $20,0($17)      # 1 1
+        ldq     $1,0($16)       # 1 1
+
+        br      $42
+
+        .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:
+        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
+        .ent bn_mul_words
+bn_mul_words:
+bn_mul_words..ng:
+        .frame $30,0,$26,0
+        .prologue 0
+        .align 5
+        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
+        .ent bn_sqr_words
+bn_sqr_words:
+bn_sqr_words..ng:
+        .frame $30,0,$26,0
+        .prologue 0
+
+        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]
+
+        addq    $16,64,$16
+        addq    $17,32,$17
+        stq     $8,-16($16)     # r[0]
+        stq     $4,-8($16)      # r[1]
+
+        blt     $18,$543
+        ldq     $20,0($17)      # 1 1
+        br      $542
+
+$442:
+        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
+$543:
+        addq    $18,4,$18
+        bgt     $18,$442        # goto tail code
+        ret     $31,($26),1     # else exit
+        .end bn_sqr_words
+
+        .align 3
+        .globl bn_add_words
+        .ent bn_add_words
+bn_add_words:
+bn_add_words..ng:
+        .frame $30,0,$26,0
+        .prologue 0
+
+        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:
+        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++
+
+        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++
+
+        bgt     $19,$945
+        ret     $31,($26),1     # else exit
+
+$900:
+        addq    $19,4,$19
+        bgt     $19,$945        # goto tail code
+        ret     $31,($26),1     # else exit
+        .end bn_add_words
+
+ #
+ # What follows was taken directly from the C compiler with a few
+ # hacks to redo the lables.
+ #
+.text
+        .align 3
+        .globl bn_div64
+        .ent bn_div64
+bn_div64:
+        ldgp $29,0($27)
+bn_div64..ng:
+        lda $30,-48($30)
+        .frame $30,48,$26,0
+        stq $26,0($30)
+        stq $9,8($30)
+        stq $10,16($30)
+        stq $11,24($30)
+        stq $12,32($30)
+        stq $13,40($30)
+        .mask 0x4003e00,-48
+        .prologue 1
+        bis $16,$16,$9
+        bis $17,$17,$10
+        bis $18,$18,$11
+        bis $31,$31,$13
+        bis $31,2,$12
+        bne $11,$119
+        lda $0,-1
+        br $31,$136
+        .align 4
+$119:
+        bis $11,$11,$16
+        jsr $26,BN_num_bits_word
+        ldgp $29,0($26)
+        subq $0,64,$1
+        beq $1,$120
+        bis $31,1,$1
+        sll $1,$0,$1
+        cmpule $9,$1,$1
+        bne $1,$120
+ #      lda $16,_IO_stderr_
+ #      lda $17,$C32
+ #      bis $0,$0,$18
+ #      jsr $26,fprintf
+ #      ldgp $29,0($26)
+        jsr $26,abort
+        ldgp $29,0($26)
+        .align 4
+$120:
+        bis $31,64,$3
+        cmpult $9,$11,$2
+        subq $3,$0,$1
+        addl $1,$31,$0
+        subq $9,$11,$1
+        cmoveq $2,$1,$9
+        beq $0,$122
+        zapnot $0,15,$2
+        subq $3,$0,$1
+        sll $11,$2,$11
+        sll $9,$2,$3
+        srl $10,$1,$1
+        sll $10,$2,$10
+        bis $3,$1,$9
+$122:
+        srl $11,32,$5
+        zapnot $11,15,$6
+        lda $7,-1
+        .align 5
+$123:
+        srl $9,32,$1
+        subq $1,$5,$1
+        bne $1,$126
+        zapnot $7,15,$27
+        br $31,$127
+        .align 4
+$126:
+        bis $9,$9,$24
+        bis $5,$5,$25
+        divqu $24,$25,$27
+$127:
+        srl $10,32,$4
+        .align 5
+$128:
+        mulq $27,$5,$1
+        subq $9,$1,$3
+        zapnot $3,240,$1
+        bne $1,$129
+        mulq $6,$27,$2
+        sll $3,32,$1
+        addq $1,$4,$1
+        cmpule $2,$1,$2
+        bne $2,$129
+        subq $27,1,$27
+        br $31,$128
+        .align 4
+$129:
+        mulq $27,$6,$1
+        mulq $27,$5,$4
+        srl $1,32,$3
+        sll $1,32,$1
+        addq $4,$3,$4
+        cmpult $10,$1,$2
+        subq $10,$1,$10
+        addq $2,$4,$2
+        cmpult $9,$2,$1
+        bis $2,$2,$4
+        beq $1,$134
+        addq $9,$11,$9
+        subq $27,1,$27
+$134:
+        subl $12,1,$12
+        subq $9,$4,$9
+        beq $12,$124
+        sll $27,32,$13
+        sll $9,32,$2
+        srl $10,32,$1
+        sll $10,32,$10
+        bis $2,$1,$9
+        br $31,$123
+        .align 4
+$124:
+        bis $13,$27,$0
+$136:
+        ldq $26,0($30)
+        ldq $9,8($30)
+        ldq $10,16($30)
+        ldq $11,24($30)
+        ldq $12,32($30)
+        ldq $13,40($30)
+        addq $30,48,$30
+        ret $31,($26),1
+        .end bn_div64
+
+        .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
diff --git a/src/lib/libcrypto/bn/asm/bn-586.pl b/src/lib/libcrypto/bn/asm/bn-586.pl
index 19d425ee96..5191bed273 100644
--- a/src/lib/libcrypto/bn/asm/bn-586.pl
+++ b/src/lib/libcrypto/bn/asm/bn-586.pl
@@ -1,18 +1,16 @@
-#!/usr/bin/perl
-#
-
 #!/usr/local/bin/perl
 
 push(@INC,"perlasm","../../perlasm");
 require "x86asm.pl";
 
-&asm_init($ARGV[0],"bn-586.pl");
+&asm_init($ARGV[0],$0);
 
 &bn_mul_add_words("bn_mul_add_words");
 &bn_mul_words("bn_mul_words");
 &bn_sqr_words("bn_sqr_words");
-&bn_div64("bn_div64");
+&bn_div_words("bn_div_words");
 &bn_add_words("bn_add_words");
+&bn_sub_words("bn_sub_words");
 
 &asm_finish();
 
@@ -228,7 +226,7 @@ sub bn_sqr_words
         &function_end($name);
         }
 
-sub bn_div64
+sub bn_div_words
         {
         local($name)=@_;
 
@@ -307,7 +305,79 @@ sub bn_add_words
                 }
         &set_label("aw_end",0);
 
-        &mov("eax",$c);
+#       &mov("eax",$c);         # $c is "eax"
+
+        &function_end($name);
+        }
+
+sub bn_sub_words
+        {
+        local($name)=@_;
+
+        &function_begin($name,"");
+
+        &comment("");
+        $a="esi";
+        $b="edi";
+        $c="eax";
+        $r="ebx";
+        $tmp1="ecx";
+        $tmp2="edx";
+        $num="ebp";
+
+        &mov($r,&wparam(0));    # get r
+         &mov($a,&wparam(1));   # get a
+        &mov($b,&wparam(2));    # get b
+         &mov($num,&wparam(3)); # get num
+        &xor($c,$c);            # clear carry
+         &and($num,0xfffffff8); # num / 8
+
+        &jz(&label("aw_finish"));
+
+        &set_label("aw_loop",0);
+        for ($i=0; $i<8; $i++)
+                {
+                &comment("Round $i");
+
+                &mov($tmp1,&DWP($i*4,$a,"",0));         # *a
+                 &mov($tmp2,&DWP($i*4,$b,"",0));        # *b
+                &sub($tmp1,$c);
+                 &mov($c,0);
+                &adc($c,$c);
+                 &sub($tmp1,$tmp2);
+                &adc($c,0);
+                 &mov(&DWP($i*4,$r,"",0),$tmp1);        # *r
+                }
+
+        &comment("");
+        &add($a,32);
+         &add($b,32);
+        &add($r,32);
+         &sub($num,8);
+        &jnz(&label("aw_loop"));
+
+        &set_label("aw_finish",0);
+        &mov($num,&wparam(3));  # get num
+        &and($num,7);
+         &jz(&label("aw_end"));
+
+        for ($i=0; $i<7; $i++)
+                {
+                &comment("Tail Round $i");
+                &mov($tmp1,&DWP($i*4,$a,"",0)); # *a
+                 &mov($tmp2,&DWP($i*4,$b,"",0));# *b
+                &sub($tmp1,$c);
+                 &mov($c,0);
+                &adc($c,$c);
+                 &sub($tmp1,$tmp2);
+                &adc($c,0);
+                 &dec($num) if ($i != 6);
+                &mov(&DWP($i*4,$r,"",0),$tmp1); # *a
+                 &jz(&label("aw_end")) if ($i != 6);
+                }
+        &set_label("aw_end",0);
+
+#       &mov("eax",$c);         # $c is "eax"
 
         &function_end($name);
         }
diff --git a/src/lib/libcrypto/bn/asm/bn-alpha.pl b/src/lib/libcrypto/bn/asm/bn-alpha.pl
new file mode 100644
index 0000000000..302edf2376
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/bn-alpha.pl
@@ -0,0 +1,571 @@
+#!/usr/local/bin/perl
+# I have this in perl so I can use more usefull register names and then convert
+# them into alpha registers.
+#
+
+$d=&data();
+$d =~ s/CC/0/g;
+$d =~ s/R1/1/g;
+$d =~ s/R2/2/g;
+$d =~ s/R3/3/g;
+$d =~ s/R4/4/g;
+$d =~ s/L1/5/g;
+$d =~ s/L2/6/g;
+$d =~ s/L3/7/g;
+$d =~ s/L4/8/g;
+$d =~ s/O1/22/g;
+$d =~ s/O2/23/g;
+$d =~ s/O3/24/g;
+$d =~ s/O4/25/g;
+$d =~ s/A1/20/g;
+$d =~ s/A2/21/g;
+$d =~ s/A3/27/g;
+$d =~ s/A4/28/g;
+if (0){
+}
+
+print $d;
+
+sub data
+        {
+        local($data)=<<'EOF';
+
+ # DEC Alpha assember
+ # 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_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:
+        .text
+        .align 3
+        .globl bn_mul_add_words
+        .ent bn_mul_add_words
+bn_mul_add_words:
+bn_mul_add_words..ng:
+        .frame $30,0,$26,0
+        .prologue 0
+        .align 5
+        subq    $18,4,$18
+        bis     $31,$31,$CC
+        blt     $18,$43         # if we are -1, -2, -3 or -4 goto tail code
+        ldq     $A1,0($17)      # 1 1
+        ldq     $R1,0($16)      # 1 1
+        .align 3
+$42:
+        mulq    $A1,$19,$L1     # 1 2 1 ######
+        ldq     $A2,8($17)      # 2 1
+        ldq     $R2,8($16)      # 2 1
+        umulh   $A1,$19,$A1     # 1 2   ######
+        ldq     $A3,16($17)     # 3 1
+        ldq     $R3,16($16)     # 3 1
+        mulq    $A2,$19,$L2     # 2 2 1 ######
+         ldq    $A4,24($17)     # 4 1
+        addq    $R1,$L1,$R1     # 1 2 2
+         ldq    $R4,24($16)     # 4 1
+        umulh   $A2,$19,$A2     # 2 2   ######
+         cmpult $R1,$L1,$O1     # 1 2 3 1
+        addq    $A1,$O1,$A1     # 1 3 1
+         addq   $R1,$CC,$R1     # 1 2 3 1
+        mulq    $A3,$19,$L3     # 3 2 1 ######
+         cmpult $R1,$CC,$CC     # 1 2 3 2
+        addq    $R2,$L2,$R2     # 2 2 2
+         addq   $A1,$CC,$CC     # 1 3 2 
+        cmpult  $R2,$L2,$O2     # 2 2 3 1
+         addq   $A2,$O2,$A2     # 2 3 1
+        umulh   $A3,$19,$A3     # 3 2   ######
+         addq   $R2,$CC,$R2     # 2 2 3 1
+        cmpult  $R2,$CC,$CC     # 2 2 3 2
+         subq   $18,4,$18
+        mulq    $A4,$19,$L4     # 4 2 1 ######
+         addq   $A2,$CC,$CC     # 2 3 2 
+        addq    $R3,$L3,$R3     # 3 2 2
+         addq   $16,32,$16
+        cmpult  $R3,$L3,$O3     # 3 2 3 1
+         stq    $R1,-32($16)    # 1 2 4
+        umulh   $A4,$19,$A4     # 4 2   ######
+         addq   $A3,$O3,$A3     # 3 3 1
+        addq    $R3,$CC,$R3     # 3 2 3 1
+         stq    $R2,-24($16)    # 2 2 4
+        cmpult  $R3,$CC,$CC     # 3 2 3 2
+         stq    $R3,-16($16)    # 3 2 4
+        addq    $R4,$L4,$R4     # 4 2 2
+         addq   $A3,$CC,$CC     # 3 3 2 
+        cmpult  $R4,$L4,$O4     # 4 2 3 1
+         addq   $17,32,$17
+        addq    $A4,$O4,$A4     # 4 3 1
+         addq   $R4,$CC,$R4     # 4 2 3 1
+        cmpult  $R4,$CC,$CC     # 4 2 3 2
+         stq    $R4,-8($16)     # 4 2 4
+        addq    $A4,$CC,$CC     # 4 3 2 
+         blt    $18,$43
+
+        ldq     $A1,0($17)      # 1 1
+        ldq     $R1,0($16)      # 1 1
+
+        br      $42
+
+        .align 4
+$45:
+        ldq     $A1,0($17)      # 4 1
+        ldq     $R1,0($16)      # 4 1
+        mulq    $A1,$19,$L1     # 4 2 1
+        subq    $18,1,$18
+        addq    $16,8,$16
+        addq    $17,8,$17
+        umulh   $A1,$19,$A1     # 4 2
+        addq    $R1,$L1,$R1     # 4 2 2
+        cmpult  $R1,$L1,$O1     # 4 2 3 1
+        addq    $A1,$O1,$A1     # 4 3 1
+        addq    $R1,$CC,$R1     # 4 2 3 1
+        cmpult  $R1,$CC,$CC     # 4 2 3 2
+        addq    $A1,$CC,$CC     # 4 3 2 
+        stq     $R1,-8($16)     # 4 2 4
+        bgt     $18,$45
+        ret     $31,($26),1     # else exit
+
+        .align 4
+$43:
+        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
+        .ent bn_mul_words
+bn_mul_words:
+bn_mul_words..ng:
+        .frame $30,0,$26,0
+        .prologue 0
+        .align 5
+        subq    $18,4,$18
+        bis     $31,$31,$CC
+        blt     $18,$143        # if we are -1, -2, -3 or -4 goto tail code
+        ldq     $A1,0($17)      # 1 1
+        .align 3
+$142:
+
+        mulq    $A1,$19,$L1     # 1 2 1 #####
+         ldq    $A2,8($17)      # 2 1
+         ldq    $A3,16($17)     # 3 1
+        umulh   $A1,$19,$A1     # 1 2   #####
+         ldq    $A4,24($17)     # 4 1
+        mulq    $A2,$19,$L2     # 2 2 1 #####
+         addq   $L1,$CC,$L1     # 1 2 3 1
+        subq    $18,4,$18
+         cmpult $L1,$CC,$CC     # 1 2 3 2
+        umulh   $A2,$19,$A2     # 2 2   #####
+         addq   $A1,$CC,$CC     # 1 3 2 
+        addq    $17,32,$17
+         addq   $L2,$CC,$L2     # 2 2 3 1
+        mulq    $A3,$19,$L3     # 3 2 1 #####
+         cmpult $L2,$CC,$CC     # 2 2 3 2
+        addq    $A2,$CC,$CC     # 2 3 2 
+         addq   $16,32,$16
+        umulh   $A3,$19,$A3     # 3 2   #####
+         stq    $L1,-32($16)    # 1 2 4
+        mulq    $A4,$19,$L4     # 4 2 1 #####
+         addq   $L3,$CC,$L3     # 3 2 3 1
+        stq     $L2,-24($16)    # 2 2 4
+         cmpult $L3,$CC,$CC     # 3 2 3 2
+        umulh   $A4,$19,$A4     # 4 2   #####
+         addq   $A3,$CC,$CC     # 3 3 2 
+        stq     $L3,-16($16)    # 3 2 4
+         addq   $L4,$CC,$L4     # 4 2 3 1
+        cmpult  $L4,$CC,$CC     # 4 2 3 2
+
+        addq    $A4,$CC,$CC     # 4 3 2 
+
+        stq     $L4,-8($16)     # 4 2 4
+
+        blt     $18,$143
+
+        ldq     $A1,0($17)      # 1 1
+
+        br      $142
+
+        .align 4
+$145:
+        ldq     $A1,0($17)      # 4 1
+        mulq    $A1,$19,$L1     # 4 2 1
+        subq    $18,1,$18
+        umulh   $A1,$19,$A1     # 4 2
+        addq    $L1,$CC,$L1     # 4 2 3 1
+         addq   $16,8,$16
+        cmpult  $L1,$CC,$CC     # 4 2 3 2
+         addq   $17,8,$17
+        addq    $A1,$CC,$CC     # 4 3 2 
+        stq     $L1,-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
+        .ent bn_sqr_words
+bn_sqr_words:
+bn_sqr_words..ng:
+        .frame $30,0,$26,0
+        .prologue 0
+
+        subq    $18,4,$18
+        blt     $18,$543        # if we are -1, -2, -3 or -4 goto tail code
+        ldq     $A1,0($17)      # 1 1
+        .align 3
+$542:
+        mulq    $A1,$A1,$L1             ######
+         ldq    $A2,8($17)      # 1 1
+        subq    $18,4
+        umulh   $A1,$A1,$R1             ######
+        ldq     $A3,16($17)     # 1 1
+        mulq    $A2,$A2,$L2             ######
+        ldq     $A4,24($17)     # 1 1
+        stq     $L1,0($16)      # r[0]
+        umulh   $A2,$A2,$R2             ######
+        stq     $R1,8($16)      # r[1]
+        mulq    $A3,$A3,$L3             ######
+        stq     $L2,16($16)     # r[0]
+        umulh   $A3,$A3,$R3             ######
+        stq     $R2,24($16)     # r[1]
+        mulq    $A4,$A4,$L4             ######
+        stq     $L3,32($16)     # r[0]
+        umulh   $A4,$A4,$R4             ######
+        stq     $R3,40($16)     # r[1]
+
+        addq    $16,64,$16
+        addq    $17,32,$17
+        stq     $L4,-16($16)    # r[0]
+        stq     $R4,-8($16)     # r[1]
+
+        blt     $18,$543
+        ldq     $A1,0($17)      # 1 1
+        br      $542
+
+$442:
+        ldq     $A1,0($17)   # a[0]
+        mulq    $A1,$A1,$L1  # a[0]*w low part       r2
+        addq    $16,16,$16
+        addq    $17,8,$17
+        subq    $18,1,$18
+        umulh   $A1,$A1,$R1  # a[0]*w high part       r3
+        stq     $L1,-16($16)   # r[0]
+        stq     $R1,-8($16)   # r[1]
+
+        bgt     $18,$442
+        ret     $31,($26),1     # else exit
+
+        .align 4
+$543:
+        addq    $18,4,$18
+        bgt     $18,$442        # goto tail code
+        ret     $31,($26),1     # else exit
+        .end bn_sqr_words
+
+        .align 3
+        .globl bn_add_words
+        .ent bn_add_words
+bn_add_words:
+bn_add_words..ng:
+        .frame $30,0,$26,0
+        .prologue 0
+
+        subq    $19,4,$19
+        bis     $31,$31,$CC     # carry = 0
+        blt     $19,$900
+        ldq     $L1,0($17)      # a[0]
+        ldq     $R1,0($18)      # b[1]
+        .align 3
+$901:
+        addq    $R1,$L1,$R1     # r=a+b;
+         ldq    $L2,8($17)      # a[1]
+        cmpult  $R1,$L1,$O1     # did we overflow?
+         ldq    $R2,8($18)      # b[1]
+        addq    $R1,$CC,$R1     # c+= overflow
+         ldq    $L3,16($17)     # a[2]
+        cmpult  $R1,$CC,$CC     # overflow?
+         ldq    $R3,16($18)     # b[2]
+        addq    $CC,$O1,$CC
+         ldq    $L4,24($17)     # a[3]
+        addq    $R2,$L2,$R2     # r=a+b;
+         ldq    $R4,24($18)     # b[3]
+        cmpult  $R2,$L2,$O2     # did we overflow?
+         addq   $R3,$L3,$R3     # r=a+b;
+        addq    $R2,$CC,$R2     # c+= overflow
+         cmpult $R3,$L3,$O3     # did we overflow?
+        cmpult  $R2,$CC,$CC     # overflow?
+         addq   $R4,$L4,$R4     # r=a+b;
+        addq    $CC,$O2,$CC
+         cmpult $R4,$L4,$O4     # did we overflow?
+        addq    $R3,$CC,$R3     # c+= overflow
+         stq    $R1,0($16)      # r[0]=c
+        cmpult  $R3,$CC,$CC     # overflow?
+         stq    $R2,8($16)      # r[1]=c
+        addq    $CC,$O3,$CC
+         stq    $R3,16($16)     # r[2]=c
+        addq    $R4,$CC,$R4     # c+= overflow
+         subq   $19,4,$19       # loop--
+        cmpult  $R4,$CC,$CC     # overflow?
+         addq   $17,32,$17      # a++
+        addq    $CC,$O4,$CC
+         stq    $R4,24($16)     # r[3]=c
+        addq    $18,32,$18      # b++
+         addq   $16,32,$16      # r++
+
+        blt     $19,$900
+         ldq    $L1,0($17)      # a[0]
+        ldq     $R1,0($18)      # b[1]
+         br     $901
+        .align 4
+$945:
+        ldq     $L1,0($17)      # a[0]
+         ldq    $R1,0($18)      # b[1]
+        addq    $R1,$L1,$R1     # r=a+b;
+         subq   $19,1,$19       # loop--
+        addq    $R1,$CC,$R1     # c+= overflow
+         addq   $17,8,$17       # a++
+        cmpult  $R1,$L1,$O1     # did we overflow?
+         cmpult $R1,$CC,$CC     # overflow?
+        addq    $18,8,$18       # b++
+         stq    $R1,0($16)      # r[0]=c
+        addq    $CC,$O1,$CC
+         addq   $16,8,$16       # r++
+
+        bgt     $19,$945
+        ret     $31,($26),1     # else exit
+
+$900:
+        addq    $19,4,$19
+        bgt     $19,$945        # goto tail code
+        ret     $31,($26),1     # else exit
+        .end bn_add_words
+
+        .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,$31,$CC     # carry = 0
+ br     $800
+        blt     $19,$800
+        ldq     $L1,0($17)      # a[0]
+        ldq     $R1,0($18)      # b[1]
+        .align 3
+$801:
+        addq    $R1,$L1,$R1     # r=a+b;
+         ldq    $L2,8($17)      # a[1]
+        cmpult  $R1,$L1,$O1     # did we overflow?
+         ldq    $R2,8($18)      # b[1]
+        addq    $R1,$CC,$R1     # c+= overflow
+         ldq    $L3,16($17)     # a[2]
+        cmpult  $R1,$CC,$CC     # overflow?
+         ldq    $R3,16($18)     # b[2]
+        addq    $CC,$O1,$CC
+         ldq    $L4,24($17)     # a[3]
+        addq    $R2,$L2,$R2     # r=a+b;
+         ldq    $R4,24($18)     # b[3]
+        cmpult  $R2,$L2,$O2     # did we overflow?
+         addq   $R3,$L3,$R3     # r=a+b;
+        addq    $R2,$CC,$R2     # c+= overflow
+         cmpult $R3,$L3,$O3     # did we overflow?
+        cmpult  $R2,$CC,$CC     # overflow?
+         addq   $R4,$L4,$R4     # r=a+b;
+        addq    $CC,$O2,$CC
+         cmpult $R4,$L4,$O4     # did we overflow?
+        addq    $R3,$CC,$R3     # c+= overflow
+         stq    $R1,0($16)      # r[0]=c
+        cmpult  $R3,$CC,$CC     # overflow?
+         stq    $R2,8($16)      # r[1]=c
+        addq    $CC,$O3,$CC
+         stq    $R3,16($16)     # r[2]=c
+        addq    $R4,$CC,$R4     # c+= overflow
+         subq   $19,4,$19       # loop--
+        cmpult  $R4,$CC,$CC     # overflow?
+         addq   $17,32,$17      # a++
+        addq    $CC,$O4,$CC
+         stq    $R4,24($16)     # r[3]=c
+        addq    $18,32,$18      # b++
+         addq   $16,32,$16      # r++
+
+        blt     $19,$800
+         ldq    $L1,0($17)      # a[0]
+        ldq     $R1,0($18)      # b[1]
+         br     $801
+        .align 4
+$845:
+        ldq     $L1,0($17)      # a[0]
+         ldq    $R1,0($18)      # b[1]
+        cmpult  $L1,$R1,$O1     # will we borrow?
+         subq   $L1,$R1,$R1     # r=a-b;
+        subq    $19,1,$19       # loop--
+         cmpult  $R1,$CC,$O2    # will we borrow?
+        subq    $R1,$CC,$R1     # c+= overflow
+         addq   $17,8,$17       # a++
+        addq    $18,8,$18       # b++
+         stq    $R1,0($16)      # r[0]=c
+        addq    $O2,$O1,$CC
+         addq   $16,8,$16       # r++
+
+        bgt     $19,$845
+        ret     $31,($26),1     # else exit
+
+$800:
+        addq    $19,4,$19
+        bgt     $19,$845        # goto tail code
+        ret     $31,($26),1     # else exit
+        .end bn_sub_words
+
+ #
+ # What follows was taken directly from the C compiler with a few
+ # hacks to redo the lables.
+ #
+.text
+        .align 3
+        .globl bn_div_words
+        .ent bn_div_words
+bn_div_words:
+        ldgp $29,0($27)
+bn_div_words..ng:
+        lda $30,-48($30)
+        .frame $30,48,$26,0
+        stq $26,0($30)
+        stq $9,8($30)
+        stq $10,16($30)
+        stq $11,24($30)
+        stq $12,32($30)
+        stq $13,40($30)
+        .mask 0x4003e00,-48
+        .prologue 1
+        bis $16,$16,$9
+        bis $17,$17,$10
+        bis $18,$18,$11
+        bis $31,$31,$13
+        bis $31,2,$12
+        bne $11,$119
+        lda $0,-1
+        br $31,$136
+        .align 4
+$119:
+        bis $11,$11,$16
+        jsr $26,BN_num_bits_word
+        ldgp $29,0($26)
+        subq $0,64,$1
+        beq $1,$120
+        bis $31,1,$1
+        sll $1,$0,$1
+        cmpule $9,$1,$1
+        bne $1,$120
+ #      lda $16,_IO_stderr_
+ #      lda $17,$C32
+ #      bis $0,$0,$18
+ #      jsr $26,fprintf
+ #      ldgp $29,0($26)
+        jsr $26,abort
+        ldgp $29,0($26)
+        .align 4
+$120:
+        bis $31,64,$3
+        cmpult $9,$11,$2
+        subq $3,$0,$1
+        addl $1,$31,$0
+        subq $9,$11,$1
+        cmoveq $2,$1,$9
+        beq $0,$122
+        zapnot $0,15,$2
+        subq $3,$0,$1
+        sll $11,$2,$11
+        sll $9,$2,$3
+        srl $10,$1,$1
+        sll $10,$2,$10
+        bis $3,$1,$9
+$122:
+        srl $11,32,$5
+        zapnot $11,15,$6
+        lda $7,-1
+        .align 5
+$123:
+        srl $9,32,$1
+        subq $1,$5,$1
+        bne $1,$126
+        zapnot $7,15,$27
+        br $31,$127
+        .align 4
+$126:
+        bis $9,$9,$24
+        bis $5,$5,$25
+        divqu $24,$25,$27
+$127:
+        srl $10,32,$4
+        .align 5
+$128:
+        mulq $27,$5,$1
+        subq $9,$1,$3
+        zapnot $3,240,$1
+        bne $1,$129
+        mulq $6,$27,$2
+        sll $3,32,$1
+        addq $1,$4,$1
+        cmpule $2,$1,$2
+        bne $2,$129
+        subq $27,1,$27
+        br $31,$128
+        .align 4
+$129:
+        mulq $27,$6,$1
+        mulq $27,$5,$4
+        srl $1,32,$3
+        sll $1,32,$1
+        addq $4,$3,$4
+        cmpult $10,$1,$2
+        subq $10,$1,$10
+        addq $2,$4,$2
+        cmpult $9,$2,$1
+        bis $2,$2,$4
+        beq $1,$134
+        addq $9,$11,$9
+        subq $27,1,$27
+$134:
+        subl $12,1,$12
+        subq $9,$4,$9
+        beq $12,$124
+        sll $27,32,$13
+        sll $9,32,$2
+        srl $10,32,$1
+        sll $10,32,$10
+        bis $2,$1,$9
+        br $31,$123
+        .align 4
+$124:
+        bis $13,$27,$0
+$136:
+        ldq $26,0($30)
+        ldq $9,8($30)
+        ldq $10,16($30)
+        ldq $11,24($30)
+        ldq $12,32($30)
+        ldq $13,40($30)
+        addq $30,48,$30
+        ret $31,($26),1
+        .end bn_div_words
+EOF
+        return($data);
+        }
+
diff --git a/src/lib/libcrypto/bn/asm/bn-win32.asm b/src/lib/libcrypto/bn/asm/bn-win32.asm
index 017ea462b0..871bd88d77 100644
--- a/src/lib/libcrypto/bn/asm/bn-win32.asm
+++ b/src/lib/libcrypto/bn/asm/bn-win32.asm
@@ -485,9 +485,9 @@ $L010sw_end:
 _bn_sqr_words ENDP
 _TEXT   ENDS
 _TEXT   SEGMENT
-PUBLIC  _bn_div64
+PUBLIC  _bn_div_words
 
-_bn_div64 PROC NEAR
+_bn_div_words PROC NEAR
         push    ebp
         push    ebx
         push    esi
@@ -501,7 +501,7 @@ _bn_div64 PROC NEAR
         pop     ebx
         pop     ebp
         ret
-_bn_div64 ENDP
+_bn_div_words ENDP
 _TEXT   ENDS
 _TEXT   SEGMENT
 PUBLIC  _bn_add_words
@@ -678,7 +678,6 @@ $L011aw_finish:
         adc     eax,            0
         mov     DWORD PTR 24[ebx],ecx
 $L013aw_end:
-        mov     eax,            eax
         pop     edi
         pop     esi
         pop     ebx
@@ -686,4 +685,1438 @@ $L013aw_end:
         ret
 _bn_add_words ENDP
 _TEXT   ENDS
+_TEXT   SEGMENT
+PUBLIC  _bn_sub_words
+
+_bn_sub_words PROC NEAR
+        push    ebp
+        push    ebx
+        push    esi
+        push    edi
+        ; 
+        mov     ebx,            DWORD PTR 20[esp]
+        mov     esi,            DWORD PTR 24[esp]
+        mov     edi,            DWORD PTR 28[esp]
+        mov     ebp,            DWORD PTR 32[esp]
+        xor     eax,            eax
+        and     ebp,            4294967288
+        jz      $L014aw_finish
+L015aw_loop:
+        ; Round 0
+        mov     ecx,            DWORD PTR [esi]
+        mov     edx,            DWORD PTR [edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        mov     DWORD PTR [ebx],ecx
+        ; Round 1
+        mov     ecx,            DWORD PTR 4[esi]
+        mov     edx,            DWORD PTR 4[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        mov     DWORD PTR 4[ebx],ecx
+        ; Round 2
+        mov     ecx,            DWORD PTR 8[esi]
+        mov     edx,            DWORD PTR 8[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        mov     DWORD PTR 8[ebx],ecx
+        ; Round 3
+        mov     ecx,            DWORD PTR 12[esi]
+        mov     edx,            DWORD PTR 12[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        mov     DWORD PTR 12[ebx],ecx
+        ; Round 4
+        mov     ecx,            DWORD PTR 16[esi]
+        mov     edx,            DWORD PTR 16[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        mov     DWORD PTR 16[ebx],ecx
+        ; Round 5
+        mov     ecx,            DWORD PTR 20[esi]
+        mov     edx,            DWORD PTR 20[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        mov     DWORD PTR 20[ebx],ecx
+        ; Round 6
+        mov     ecx,            DWORD PTR 24[esi]
+        mov     edx,            DWORD PTR 24[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        mov     DWORD PTR 24[ebx],ecx
+        ; Round 7
+        mov     ecx,            DWORD PTR 28[esi]
+        mov     edx,            DWORD PTR 28[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        mov     DWORD PTR 28[ebx],ecx
+        ; 
+        add     esi,            32
+        add     edi,            32
+        add     ebx,            32
+        sub     ebp,            8
+        jnz     L015aw_loop
+$L014aw_finish:
+        mov     ebp,            DWORD PTR 32[esp]
+        and     ebp,            7
+        jz      $L016aw_end
+        ; Tail Round 0
+        mov     ecx,            DWORD PTR [esi]
+        mov     edx,            DWORD PTR [edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        dec     ebp
+        mov     DWORD PTR [ebx],ecx
+        jz      $L016aw_end
+        ; Tail Round 1
+        mov     ecx,            DWORD PTR 4[esi]
+        mov     edx,            DWORD PTR 4[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        dec     ebp
+        mov     DWORD PTR 4[ebx],ecx
+        jz      $L016aw_end
+        ; Tail Round 2
+        mov     ecx,            DWORD PTR 8[esi]
+        mov     edx,            DWORD PTR 8[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        dec     ebp
+        mov     DWORD PTR 8[ebx],ecx
+        jz      $L016aw_end
+        ; Tail Round 3
+        mov     ecx,            DWORD PTR 12[esi]
+        mov     edx,            DWORD PTR 12[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        dec     ebp
+        mov     DWORD PTR 12[ebx],ecx
+        jz      $L016aw_end
+        ; Tail Round 4
+        mov     ecx,            DWORD PTR 16[esi]
+        mov     edx,            DWORD PTR 16[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        dec     ebp
+        mov     DWORD PTR 16[ebx],ecx
+        jz      $L016aw_end
+        ; Tail Round 5
+        mov     ecx,            DWORD PTR 20[esi]
+        mov     edx,            DWORD PTR 20[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        dec     ebp
+        mov     DWORD PTR 20[ebx],ecx
+        jz      $L016aw_end
+        ; Tail Round 6
+        mov     ecx,            DWORD PTR 24[esi]
+        mov     edx,            DWORD PTR 24[edi]
+        sub     ecx,            eax
+        mov     eax,            0
+        adc     eax,            eax
+        sub     ecx,            edx
+        adc     eax,            0
+        mov     DWORD PTR 24[ebx],ecx
+$L016aw_end:
+        pop     edi
+        pop     esi
+        pop     ebx
+        pop     ebp
+        ret
+_bn_sub_words ENDP
+_TEXT   ENDS
+_TEXT   SEGMENT
+PUBLIC  _bn_mul_comba8
+
+_bn_mul_comba8 PROC NEAR
+        push    esi
+        mov     esi,            DWORD PTR 12[esp]
+        push    edi
+        mov     edi,            DWORD PTR 20[esp]
+        push    ebp
+        push    ebx
+        xor     ebx,            ebx
+        mov     eax,            DWORD PTR [esi]
+        xor     ecx,            ecx
+        mov     edx,            DWORD PTR [edi]
+        ; ################## Calculate word 0
+        xor     ebp,            ebp
+        ; mul a[0]*b[0]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR [edi]
+        adc     ebp,            0
+        mov     DWORD PTR [eax],ebx
+        mov     eax,            DWORD PTR 4[esi]
+        ; saved r[0]
+        ; ################## Calculate word 1
+        xor     ebx,            ebx
+        ; mul a[1]*b[0]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR [esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ebx,            0
+        ; mul a[0]*b[1]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR [edi]
+        adc     ebx,            0
+        mov     DWORD PTR 4[eax],ecx
+        mov     eax,            DWORD PTR 8[esi]
+        ; saved r[1]
+        ; ################## Calculate word 2
+        xor     ecx,            ecx
+        ; mul a[2]*b[0]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ecx,            0
+        ; mul a[1]*b[1]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR [esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ecx,            0
+        ; mul a[0]*b[2]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR [edi]
+        adc     ecx,            0
+        mov     DWORD PTR 8[eax],ebp
+        mov     eax,            DWORD PTR 12[esi]
+        ; saved r[2]
+        ; ################## Calculate word 3
+        xor     ebp,            ebp
+        ; mul a[3]*b[0]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ebp,            0
+        ; mul a[2]*b[1]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ebp,            0
+        ; mul a[1]*b[2]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR [esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ebp,            0
+        ; mul a[0]*b[3]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR [edi]
+        adc     ebp,            0
+        mov     DWORD PTR 12[eax],ebx
+        mov     eax,            DWORD PTR 16[esi]
+        ; saved r[3]
+        ; ################## Calculate word 4
+        xor     ebx,            ebx
+        ; mul a[4]*b[0]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ebx,            0
+        ; mul a[3]*b[1]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ebx,            0
+        ; mul a[2]*b[2]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ebx,            0
+        ; mul a[1]*b[3]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR [esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 16[edi]
+        adc     ebx,            0
+        ; mul a[0]*b[4]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR [edi]
+        adc     ebx,            0
+        mov     DWORD PTR 16[eax],ecx
+        mov     eax,            DWORD PTR 20[esi]
+        ; saved r[4]
+        ; ################## Calculate word 5
+        xor     ecx,            ecx
+        ; mul a[5]*b[0]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ecx,            0
+        ; mul a[4]*b[1]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ecx,            0
+        ; mul a[3]*b[2]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ecx,            0
+        ; mul a[2]*b[3]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 16[edi]
+        adc     ecx,            0
+        ; mul a[1]*b[4]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR [esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 20[edi]
+        adc     ecx,            0
+        ; mul a[0]*b[5]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR [edi]
+        adc     ecx,            0
+        mov     DWORD PTR 20[eax],ebp
+        mov     eax,            DWORD PTR 24[esi]
+        ; saved r[5]
+        ; ################## Calculate word 6
+        xor     ebp,            ebp
+        ; mul a[6]*b[0]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ebp,            0
+        ; mul a[5]*b[1]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ebp,            0
+        ; mul a[4]*b[2]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ebp,            0
+        ; mul a[3]*b[3]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 16[edi]
+        adc     ebp,            0
+        ; mul a[2]*b[4]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 20[edi]
+        adc     ebp,            0
+        ; mul a[1]*b[5]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR [esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 24[edi]
+        adc     ebp,            0
+        ; mul a[0]*b[6]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR [edi]
+        adc     ebp,            0
+        mov     DWORD PTR 24[eax],ebx
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[6]
+        ; ################## Calculate word 7
+        xor     ebx,            ebx
+        ; mul a[7]*b[0]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ebx,            0
+        ; mul a[6]*b[1]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ebx,            0
+        ; mul a[5]*b[2]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ebx,            0
+        ; mul a[4]*b[3]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 16[edi]
+        adc     ebx,            0
+        ; mul a[3]*b[4]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 20[edi]
+        adc     ebx,            0
+        ; mul a[2]*b[5]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 24[edi]
+        adc     ebx,            0
+        ; mul a[1]*b[6]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR [esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 28[edi]
+        adc     ebx,            0
+        ; mul a[0]*b[7]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ebx,            0
+        mov     DWORD PTR 28[eax],ecx
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[7]
+        ; ################## Calculate word 8
+        xor     ecx,            ecx
+        ; mul a[7]*b[1]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ecx,            0
+        ; mul a[6]*b[2]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ecx,            0
+        ; mul a[5]*b[3]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 16[edi]
+        adc     ecx,            0
+        ; mul a[4]*b[4]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 20[edi]
+        adc     ecx,            0
+        ; mul a[3]*b[5]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 24[edi]
+        adc     ecx,            0
+        ; mul a[2]*b[6]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 28[edi]
+        adc     ecx,            0
+        ; mul a[1]*b[7]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ecx,            0
+        mov     DWORD PTR 32[eax],ebp
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[8]
+        ; ################## Calculate word 9
+        xor     ebp,            ebp
+        ; mul a[7]*b[2]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ebp,            0
+        ; mul a[6]*b[3]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 16[edi]
+        adc     ebp,            0
+        ; mul a[5]*b[4]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 20[edi]
+        adc     ebp,            0
+        ; mul a[4]*b[5]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 24[edi]
+        adc     ebp,            0
+        ; mul a[3]*b[6]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 28[edi]
+        adc     ebp,            0
+        ; mul a[2]*b[7]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ebp,            0
+        mov     DWORD PTR 36[eax],ebx
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[9]
+        ; ################## Calculate word 10
+        xor     ebx,            ebx
+        ; mul a[7]*b[3]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 16[edi]
+        adc     ebx,            0
+        ; mul a[6]*b[4]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 20[edi]
+        adc     ebx,            0
+        ; mul a[5]*b[5]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 24[edi]
+        adc     ebx,            0
+        ; mul a[4]*b[6]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 28[edi]
+        adc     ebx,            0
+        ; mul a[3]*b[7]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 16[edi]
+        adc     ebx,            0
+        mov     DWORD PTR 40[eax],ecx
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[10]
+        ; ################## Calculate word 11
+        xor     ecx,            ecx
+        ; mul a[7]*b[4]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 20[edi]
+        adc     ecx,            0
+        ; mul a[6]*b[5]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 24[edi]
+        adc     ecx,            0
+        ; mul a[5]*b[6]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 28[edi]
+        adc     ecx,            0
+        ; mul a[4]*b[7]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 20[edi]
+        adc     ecx,            0
+        mov     DWORD PTR 44[eax],ebp
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[11]
+        ; ################## Calculate word 12
+        xor     ebp,            ebp
+        ; mul a[7]*b[5]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 24[edi]
+        adc     ebp,            0
+        ; mul a[6]*b[6]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 28[edi]
+        adc     ebp,            0
+        ; mul a[5]*b[7]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 24[edi]
+        adc     ebp,            0
+        mov     DWORD PTR 48[eax],ebx
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[12]
+        ; ################## Calculate word 13
+        xor     ebx,            ebx
+        ; mul a[7]*b[6]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 28[edi]
+        adc     ebx,            0
+        ; mul a[6]*b[7]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 28[edi]
+        adc     ebx,            0
+        mov     DWORD PTR 52[eax],ecx
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[13]
+        ; ################## Calculate word 14
+        xor     ecx,            ecx
+        ; mul a[7]*b[7]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebx,            edx
+        adc     ecx,            0
+        mov     DWORD PTR 56[eax],ebp
+        ; saved r[14]
+        ; save r[15]
+        mov     DWORD PTR 60[eax],ebx
+        pop     ebx
+        pop     ebp
+        pop     edi
+        pop     esi
+        ret
+_bn_mul_comba8 ENDP
+_TEXT   ENDS
+_TEXT   SEGMENT
+PUBLIC  _bn_mul_comba4
+
+_bn_mul_comba4 PROC NEAR
+        push    esi
+        mov     esi,            DWORD PTR 12[esp]
+        push    edi
+        mov     edi,            DWORD PTR 20[esp]
+        push    ebp
+        push    ebx
+        xor     ebx,            ebx
+        mov     eax,            DWORD PTR [esi]
+        xor     ecx,            ecx
+        mov     edx,            DWORD PTR [edi]
+        ; ################## Calculate word 0
+        xor     ebp,            ebp
+        ; mul a[0]*b[0]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR [edi]
+        adc     ebp,            0
+        mov     DWORD PTR [eax],ebx
+        mov     eax,            DWORD PTR 4[esi]
+        ; saved r[0]
+        ; ################## Calculate word 1
+        xor     ebx,            ebx
+        ; mul a[1]*b[0]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR [esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ebx,            0
+        ; mul a[0]*b[1]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR [edi]
+        adc     ebx,            0
+        mov     DWORD PTR 4[eax],ecx
+        mov     eax,            DWORD PTR 8[esi]
+        ; saved r[1]
+        ; ################## Calculate word 2
+        xor     ecx,            ecx
+        ; mul a[2]*b[0]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ecx,            0
+        ; mul a[1]*b[1]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR [esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ecx,            0
+        ; mul a[0]*b[2]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR [edi]
+        adc     ecx,            0
+        mov     DWORD PTR 8[eax],ebp
+        mov     eax,            DWORD PTR 12[esi]
+        ; saved r[2]
+        ; ################## Calculate word 3
+        xor     ebp,            ebp
+        ; mul a[3]*b[0]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ebp,            0
+        ; mul a[2]*b[1]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ebp,            0
+        ; mul a[1]*b[2]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR [esi]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ebp,            0
+        ; mul a[0]*b[3]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 4[edi]
+        adc     ebp,            0
+        mov     DWORD PTR 12[eax],ebx
+        mov     eax,            DWORD PTR 12[esi]
+        ; saved r[3]
+        ; ################## Calculate word 4
+        xor     ebx,            ebx
+        ; mul a[3]*b[1]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ebx,            0
+        ; mul a[2]*b[2]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ebx,            0
+        ; mul a[1]*b[3]
+        mul     edx
+        add     ecx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 8[edi]
+        adc     ebx,            0
+        mov     DWORD PTR 16[eax],ecx
+        mov     eax,            DWORD PTR 12[esi]
+        ; saved r[4]
+        ; ################## Calculate word 5
+        xor     ecx,            ecx
+        ; mul a[3]*b[2]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ecx,            0
+        ; mul a[2]*b[3]
+        mul     edx
+        add     ebp,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 12[edi]
+        adc     ecx,            0
+        mov     DWORD PTR 20[eax],ebp
+        mov     eax,            DWORD PTR 12[esi]
+        ; saved r[5]
+        ; ################## Calculate word 6
+        xor     ebp,            ebp
+        ; mul a[3]*b[3]
+        mul     edx
+        add     ebx,            eax
+        mov     eax,            DWORD PTR 20[esp]
+        adc     ecx,            edx
+        adc     ebp,            0
+        mov     DWORD PTR 24[eax],ebx
+        ; saved r[6]
+        ; save r[7]
+        mov     DWORD PTR 28[eax],ecx
+        pop     ebx
+        pop     ebp
+        pop     edi
+        pop     esi
+        ret
+_bn_mul_comba4 ENDP
+_TEXT   ENDS
+_TEXT   SEGMENT
+PUBLIC  _bn_sqr_comba8
+
+_bn_sqr_comba8 PROC NEAR
+        push    esi
+        push    edi
+        push    ebp
+        push    ebx
+        mov     edi,            DWORD PTR 20[esp]
+        mov     esi,            DWORD PTR 24[esp]
+        xor     ebx,            ebx
+        xor     ecx,            ecx
+        mov     eax,            DWORD PTR [esi]
+        ; ############### Calculate word 0
+        xor     ebp,            ebp
+        ; sqr a[0]*a[0]
+        mul     eax
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR [esi]
+        adc     ebp,            0
+        mov     DWORD PTR [edi],ebx
+        mov     eax,            DWORD PTR 4[esi]
+        ; saved r[0]
+        ; ############### Calculate word 1
+        xor     ebx,            ebx
+        ; sqr a[1]*a[0]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebx,            0
+        mov     DWORD PTR 4[edi],ecx
+        mov     edx,            DWORD PTR [esi]
+        ; saved r[1]
+        ; ############### Calculate word 2
+        xor     ecx,            ecx
+        ; sqr a[2]*a[0]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ecx,            0
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ecx,            0
+        ; sqr a[1]*a[1]
+        mul     eax
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR [esi]
+        adc     ecx,            0
+        mov     DWORD PTR 8[edi],ebp
+        mov     eax,            DWORD PTR 12[esi]
+        ; saved r[2]
+        ; ############### Calculate word 3
+        xor     ebp,            ebp
+        ; sqr a[3]*a[0]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebp,            0
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebp,            0
+        mov     edx,            DWORD PTR 4[esi]
+        ; sqr a[2]*a[1]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebp,            0
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ebp,            0
+        mov     DWORD PTR 12[edi],ebx
+        mov     edx,            DWORD PTR [esi]
+        ; saved r[3]
+        ; ############### Calculate word 4
+        xor     ebx,            ebx
+        ; sqr a[4]*a[0]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ebx,            0
+        mov     edx,            DWORD PTR 4[esi]
+        ; sqr a[3]*a[1]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebx,            0
+        ; sqr a[2]*a[2]
+        mul     eax
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR [esi]
+        adc     ebx,            0
+        mov     DWORD PTR 16[edi],ecx
+        mov     eax,            DWORD PTR 20[esi]
+        ; saved r[4]
+        ; ############### Calculate word 5
+        xor     ecx,            ecx
+        ; sqr a[5]*a[0]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ecx,            0
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ecx,            0
+        mov     edx,            DWORD PTR 4[esi]
+        ; sqr a[4]*a[1]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ecx,            0
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ecx,            0
+        mov     edx,            DWORD PTR 8[esi]
+        ; sqr a[3]*a[2]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ecx,            0
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ecx,            0
+        mov     DWORD PTR 20[edi],ebp
+        mov     edx,            DWORD PTR [esi]
+        ; saved r[5]
+        ; ############### Calculate word 6
+        xor     ebp,            ebp
+        ; sqr a[6]*a[0]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebp,            0
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ebp,            0
+        mov     edx,            DWORD PTR 4[esi]
+        ; sqr a[5]*a[1]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebp,            0
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ebp,            0
+        mov     edx,            DWORD PTR 8[esi]
+        ; sqr a[4]*a[2]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebp,            0
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ebp,            0
+        ; sqr a[3]*a[3]
+        mul     eax
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR [esi]
+        adc     ebp,            0
+        mov     DWORD PTR 24[edi],ebx
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[6]
+        ; ############### Calculate word 7
+        xor     ebx,            ebx
+        ; sqr a[7]*a[0]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ebx,            0
+        mov     edx,            DWORD PTR 4[esi]
+        ; sqr a[6]*a[1]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ebx,            0
+        mov     edx,            DWORD PTR 8[esi]
+        ; sqr a[5]*a[2]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ebx,            0
+        mov     edx,            DWORD PTR 12[esi]
+        ; sqr a[4]*a[3]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 28[esi]
+        adc     ebx,            0
+        mov     DWORD PTR 28[edi],ecx
+        mov     edx,            DWORD PTR 4[esi]
+        ; saved r[7]
+        ; ############### Calculate word 8
+        xor     ecx,            ecx
+        ; sqr a[7]*a[1]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ecx,            0
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ecx,            0
+        mov     edx,            DWORD PTR 8[esi]
+        ; sqr a[6]*a[2]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ecx,            0
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ecx,            0
+        mov     edx,            DWORD PTR 12[esi]
+        ; sqr a[5]*a[3]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ecx,            0
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     eax,            DWORD PTR 16[esi]
+        adc     ecx,            0
+        ; sqr a[4]*a[4]
+        mul     eax
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR 8[esi]
+        adc     ecx,            0
+        mov     DWORD PTR 32[edi],ebp
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[8]
+        ; ############### Calculate word 9
+        xor     ebp,            ebp
+        ; sqr a[7]*a[2]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebp,            0
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ebp,            0
+        mov     edx,            DWORD PTR 12[esi]
+        ; sqr a[6]*a[3]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebp,            0
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ebp,            0
+        mov     edx,            DWORD PTR 16[esi]
+        ; sqr a[5]*a[4]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebp,            0
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     eax,            DWORD PTR 28[esi]
+        adc     ebp,            0
+        mov     DWORD PTR 36[edi],ebx
+        mov     edx,            DWORD PTR 12[esi]
+        ; saved r[9]
+        ; ############### Calculate word 10
+        xor     ebx,            ebx
+        ; sqr a[7]*a[3]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ebx,            0
+        mov     edx,            DWORD PTR 16[esi]
+        ; sqr a[6]*a[4]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 20[esi]
+        adc     ebx,            0
+        ; sqr a[5]*a[5]
+        mul     eax
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 16[esi]
+        adc     ebx,            0
+        mov     DWORD PTR 40[edi],ecx
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[10]
+        ; ############### Calculate word 11
+        xor     ecx,            ecx
+        ; sqr a[7]*a[4]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ecx,            0
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ecx,            0
+        mov     edx,            DWORD PTR 20[esi]
+        ; sqr a[6]*a[5]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ecx,            0
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     eax,            DWORD PTR 28[esi]
+        adc     ecx,            0
+        mov     DWORD PTR 44[edi],ebp
+        mov     edx,            DWORD PTR 20[esi]
+        ; saved r[11]
+        ; ############### Calculate word 12
+        xor     ebp,            ebp
+        ; sqr a[7]*a[5]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebp,            0
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     eax,            DWORD PTR 24[esi]
+        adc     ebp,            0
+        ; sqr a[6]*a[6]
+        mul     eax
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR 24[esi]
+        adc     ebp,            0
+        mov     DWORD PTR 48[edi],ebx
+        mov     eax,            DWORD PTR 28[esi]
+        ; saved r[12]
+        ; ############### Calculate word 13
+        xor     ebx,            ebx
+        ; sqr a[7]*a[6]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 28[esi]
+        adc     ebx,            0
+        mov     DWORD PTR 52[edi],ecx
+        ; saved r[13]
+        ; ############### Calculate word 14
+        xor     ecx,            ecx
+        ; sqr a[7]*a[7]
+        mul     eax
+        add     ebp,            eax
+        adc     ebx,            edx
+        adc     ecx,            0
+        mov     DWORD PTR 56[edi],ebp
+        ; saved r[14]
+        mov     DWORD PTR 60[edi],ebx
+        pop     ebx
+        pop     ebp
+        pop     edi
+        pop     esi
+        ret
+_bn_sqr_comba8 ENDP
+_TEXT   ENDS
+_TEXT   SEGMENT
+PUBLIC  _bn_sqr_comba4
+
+_bn_sqr_comba4 PROC NEAR
+        push    esi
+        push    edi
+        push    ebp
+        push    ebx
+        mov     edi,            DWORD PTR 20[esp]
+        mov     esi,            DWORD PTR 24[esp]
+        xor     ebx,            ebx
+        xor     ecx,            ecx
+        mov     eax,            DWORD PTR [esi]
+        ; ############### Calculate word 0
+        xor     ebp,            ebp
+        ; sqr a[0]*a[0]
+        mul     eax
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     edx,            DWORD PTR [esi]
+        adc     ebp,            0
+        mov     DWORD PTR [edi],ebx
+        mov     eax,            DWORD PTR 4[esi]
+        ; saved r[0]
+        ; ############### Calculate word 1
+        xor     ebx,            ebx
+        ; sqr a[1]*a[0]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebx,            0
+        mov     DWORD PTR 4[edi],ecx
+        mov     edx,            DWORD PTR [esi]
+        ; saved r[1]
+        ; ############### Calculate word 2
+        xor     ecx,            ecx
+        ; sqr a[2]*a[0]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ecx,            0
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     eax,            DWORD PTR 4[esi]
+        adc     ecx,            0
+        ; sqr a[1]*a[1]
+        mul     eax
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     edx,            DWORD PTR [esi]
+        adc     ecx,            0
+        mov     DWORD PTR 8[edi],ebp
+        mov     eax,            DWORD PTR 12[esi]
+        ; saved r[2]
+        ; ############### Calculate word 3
+        xor     ebp,            ebp
+        ; sqr a[3]*a[0]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebp,            0
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebp,            0
+        mov     edx,            DWORD PTR 4[esi]
+        ; sqr a[2]*a[1]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebp,            0
+        add     ebx,            eax
+        adc     ecx,            edx
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ebp,            0
+        mov     DWORD PTR 12[edi],ebx
+        mov     edx,            DWORD PTR 4[esi]
+        ; saved r[3]
+        ; ############### Calculate word 4
+        xor     ebx,            ebx
+        ; sqr a[3]*a[1]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ebx,            0
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     eax,            DWORD PTR 8[esi]
+        adc     ebx,            0
+        ; sqr a[2]*a[2]
+        mul     eax
+        add     ecx,            eax
+        adc     ebp,            edx
+        mov     edx,            DWORD PTR 8[esi]
+        adc     ebx,            0
+        mov     DWORD PTR 16[edi],ecx
+        mov     eax,            DWORD PTR 12[esi]
+        ; saved r[4]
+        ; ############### Calculate word 5
+        xor     ecx,            ecx
+        ; sqr a[3]*a[2]
+        mul     edx
+        add     eax,            eax
+        adc     edx,            edx
+        adc     ecx,            0
+        add     ebp,            eax
+        adc     ebx,            edx
+        mov     eax,            DWORD PTR 12[esi]
+        adc     ecx,            0
+        mov     DWORD PTR 20[edi],ebp
+        ; saved r[5]
+        ; ############### Calculate word 6
+        xor     ebp,            ebp
+        ; sqr a[3]*a[3]
+        mul     eax
+        add     ebx,            eax
+        adc     ecx,            edx
+        adc     ebp,            0
+        mov     DWORD PTR 24[edi],ebx
+        ; saved r[6]
+        mov     DWORD PTR 28[edi],ecx
+        pop     ebx
+        pop     ebp
+        pop     edi
+        pop     esi
+        ret
+_bn_sqr_comba4 ENDP
+_TEXT   ENDS
 END
diff --git a/src/lib/libcrypto/bn/asm/bn86unix.cpp b/src/lib/libcrypto/bn/asm/bn86unix.cpp
deleted file mode 100644
index 64702201ea..0000000000
--- a/src/lib/libcrypto/bn/asm/bn86unix.cpp
+++ /dev/null
@@ -1,752 +0,0 @@
-/* Run the C pre-processor over this file with one of the following defined
- * ELF - elf object files,
- * OUT - a.out object files,
- * BSDI - BSDI style a.out object files
- * SOL - Solaris style elf
- */
-
-#define TYPE(a,b)       .type   a,b
-#define SIZE(a,b)       .size   a,b
-
-#if defined(OUT) || defined(BSDI)
-#define bn_mul_add_words _bn_mul_add_words
-#define bn_mul_words _bn_mul_words
-#define bn_sqr_words _bn_sqr_words
-#define bn_div64 _bn_div64
-#define bn_add_words _bn_add_words
-
-#endif
-
-#ifdef OUT
-#define OK      1
-#define ALIGN   4
-#endif
-
-#ifdef BSDI
-#define OK              1
-#define ALIGN           4
-#undef SIZE
-#undef TYPE
-#define SIZE(a,b)
-#define TYPE(a,b)
-#endif
-
-#if defined(ELF) || defined(SOL)
-#define OK              1
-#define ALIGN           16
-#endif
-
-#ifndef OK
-You need to define one of
-ELF - elf systems - linux-elf, NetBSD and DG-UX
-OUT - a.out systems - linux-a.out and FreeBSD
-SOL - solaris systems, which are elf with strange comment lines
-BSDI - a.out with a very primative version of as.
-#endif
-
-/* Let the Assembler begin :-) */
-        /* Don't even think of reading this code */
-        /* It was automatically generated by bn-586.pl */
-        /* Which is a perl program used to generate the x86 assember for */
-        /* any of elf, a.out, BSDI,Win32, or Solaris */
-        /* eric <eay@cryptsoft.com> */
-
-        .file   "bn-586.s"
-        .version        "01.01"
-gcc2_compiled.:
-.text
-        .align ALIGN
-.globl bn_mul_add_words
-        TYPE(bn_mul_add_words,@function)
-bn_mul_add_words:
-        pushl   %ebp
-        pushl   %ebx
-        pushl   %esi
-        pushl   %edi
-
-
-        xorl    %esi,           %esi
-        movl    20(%esp),       %edi
-        movl    28(%esp),       %ecx
-        movl    24(%esp),       %ebx
-        andl    $4294967288,    %ecx
-        movl    32(%esp),       %ebp
-        pushl   %ecx
-        jz      .L000maw_finish
-.L001maw_loop:
-        movl    %ecx,           (%esp)
-        /* Round 0 */
-        movl    (%ebx),         %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    (%edi),         %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           (%edi)
-        movl    %edx,           %esi
-        /* Round 4 */
-        movl    4(%ebx),        %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    4(%edi),        %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           4(%edi)
-        movl    %edx,           %esi
-        /* Round 8 */
-        movl    8(%ebx),        %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    8(%edi),        %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           8(%edi)
-        movl    %edx,           %esi
-        /* Round 12 */
-        movl    12(%ebx),       %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    12(%edi),       %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           12(%edi)
-        movl    %edx,           %esi
-        /* Round 16 */
-        movl    16(%ebx),       %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    16(%edi),       %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           16(%edi)
-        movl    %edx,           %esi
-        /* Round 20 */
-        movl    20(%ebx),       %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    20(%edi),       %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           20(%edi)
-        movl    %edx,           %esi
-        /* Round 24 */
-        movl    24(%ebx),       %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    24(%edi),       %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           24(%edi)
-        movl    %edx,           %esi
-        /* Round 28 */
-        movl    28(%ebx),       %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    28(%edi),       %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           28(%edi)
-        movl    %edx,           %esi
-
-        movl    (%esp),         %ecx
-        addl    $32,            %ebx
-        addl    $32,            %edi
-        subl    $8,             %ecx
-        jnz     .L001maw_loop
-.L000maw_finish:
-        movl    32(%esp),       %ecx
-        andl    $7,             %ecx
-        jnz     .L002maw_finish2
-        jmp     .L003maw_end
-.align ALIGN
-.L002maw_finish2:
-        /* Tail Round 0 */
-        movl    (%ebx),         %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    (%edi),         %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        decl    %ecx
-        movl    %eax,           (%edi)
-        movl    %edx,           %esi
-        jz      .L003maw_end
-        /* Tail Round 1 */
-        movl    4(%ebx),        %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    4(%edi),        %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        decl    %ecx
-        movl    %eax,           4(%edi)
-        movl    %edx,           %esi
-        jz      .L003maw_end
-        /* Tail Round 2 */
-        movl    8(%ebx),        %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    8(%edi),        %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        decl    %ecx
-        movl    %eax,           8(%edi)
-        movl    %edx,           %esi
-        jz      .L003maw_end
-        /* Tail Round 3 */
-        movl    12(%ebx),       %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    12(%edi),       %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        decl    %ecx
-        movl    %eax,           12(%edi)
-        movl    %edx,           %esi
-        jz      .L003maw_end
-        /* Tail Round 4 */
-        movl    16(%ebx),       %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    16(%edi),       %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        decl    %ecx
-        movl    %eax,           16(%edi)
-        movl    %edx,           %esi
-        jz      .L003maw_end
-        /* Tail Round 5 */
-        movl    20(%ebx),       %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    20(%edi),       %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        decl    %ecx
-        movl    %eax,           20(%edi)
-        movl    %edx,           %esi
-        jz      .L003maw_end
-        /* Tail Round 6 */
-        movl    24(%ebx),       %eax
-        mull    %ebp
-        addl    %esi,           %eax
-        movl    24(%edi),       %esi
-        adcl    $0,             %edx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           24(%edi)
-        movl    %edx,           %esi
-.L003maw_end:
-        movl    %esi,           %eax
-        popl    %ecx
-        popl    %edi
-        popl    %esi
-        popl    %ebx
-        popl    %ebp
-        ret
-.bn_mul_add_words_end:
-        SIZE(bn_mul_add_words,.bn_mul_add_words_end-bn_mul_add_words)
-.ident  "bn_mul_add_words"
-.text
-        .align ALIGN
-.globl bn_mul_words
-        TYPE(bn_mul_words,@function)
-bn_mul_words:
-        pushl   %ebp
-        pushl   %ebx
-        pushl   %esi
-        pushl   %edi
-
-
-        xorl    %esi,           %esi
-        movl    20(%esp),       %edi
-        movl    24(%esp),       %ebx
-        movl    28(%esp),       %ebp
-        movl    32(%esp),       %ecx
-        andl    $4294967288,    %ebp
-        jz      .L004mw_finish
-.L005mw_loop:
-        /* Round 0 */
-        movl    (%ebx),         %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           (%edi)
-        movl    %edx,           %esi
-        /* Round 4 */
-        movl    4(%ebx),        %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           4(%edi)
-        movl    %edx,           %esi
-        /* Round 8 */
-        movl    8(%ebx),        %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           8(%edi)
-        movl    %edx,           %esi
-        /* Round 12 */
-        movl    12(%ebx),       %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           12(%edi)
-        movl    %edx,           %esi
-        /* Round 16 */
-        movl    16(%ebx),       %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           16(%edi)
-        movl    %edx,           %esi
-        /* Round 20 */
-        movl    20(%ebx),       %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           20(%edi)
-        movl    %edx,           %esi
-        /* Round 24 */
-        movl    24(%ebx),       %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           24(%edi)
-        movl    %edx,           %esi
-        /* Round 28 */
-        movl    28(%ebx),       %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           28(%edi)
-        movl    %edx,           %esi
-
-        addl    $32,            %ebx
-        addl    $32,            %edi
-        subl    $8,             %ebp
-        jz      .L004mw_finish
-        jmp     .L005mw_loop
-.L004mw_finish:
-        movl    28(%esp),       %ebp
-        andl    $7,             %ebp
-        jnz     .L006mw_finish2
-        jmp     .L007mw_end
-.align ALIGN
-.L006mw_finish2:
-        /* Tail Round 0 */
-        movl    (%ebx),         %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           (%edi)
-        movl    %edx,           %esi
-        decl    %ebp
-        jz      .L007mw_end
-        /* Tail Round 1 */
-        movl    4(%ebx),        %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           4(%edi)
-        movl    %edx,           %esi
-        decl    %ebp
-        jz      .L007mw_end
-        /* Tail Round 2 */
-        movl    8(%ebx),        %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           8(%edi)
-        movl    %edx,           %esi
-        decl    %ebp
-        jz      .L007mw_end
-        /* Tail Round 3 */
-        movl    12(%ebx),       %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           12(%edi)
-        movl    %edx,           %esi
-        decl    %ebp
-        jz      .L007mw_end
-        /* Tail Round 4 */
-        movl    16(%ebx),       %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           16(%edi)
-        movl    %edx,           %esi
-        decl    %ebp
-        jz      .L007mw_end
-        /* Tail Round 5 */
-        movl    20(%ebx),       %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           20(%edi)
-        movl    %edx,           %esi
-        decl    %ebp
-        jz      .L007mw_end
-        /* Tail Round 6 */
-        movl    24(%ebx),       %eax
-        mull    %ecx
-        addl    %esi,           %eax
-        adcl    $0,             %edx
-        movl    %eax,           24(%edi)
-        movl    %edx,           %esi
-.L007mw_end:
-        movl    %esi,           %eax
-        popl    %edi
-        popl    %esi
-        popl    %ebx
-        popl    %ebp
-        ret
-.bn_mul_words_end:
-        SIZE(bn_mul_words,.bn_mul_words_end-bn_mul_words)
-.ident  "bn_mul_words"
-.text
-        .align ALIGN
-.globl bn_sqr_words
-        TYPE(bn_sqr_words,@function)
-bn_sqr_words:
-        pushl   %ebp
-        pushl   %ebx
-        pushl   %esi
-        pushl   %edi
-
-
-        movl    20(%esp),       %esi
-        movl    24(%esp),       %edi
-        movl    28(%esp),       %ebx
-        andl    $4294967288,    %ebx
-        jz      .L008sw_finish
-.L009sw_loop:
-        /* Round 0 */
-        movl    (%edi),         %eax
-        mull    %eax
-        movl    %eax,           (%esi)
-        movl    %edx,           4(%esi)
-        /* Round 4 */
-        movl    4(%edi),        %eax
-        mull    %eax
-        movl    %eax,           8(%esi)
-        movl    %edx,           12(%esi)
-        /* Round 8 */
-        movl    8(%edi),        %eax
-        mull    %eax
-        movl    %eax,           16(%esi)
-        movl    %edx,           20(%esi)
-        /* Round 12 */
-        movl    12(%edi),       %eax
-        mull    %eax
-        movl    %eax,           24(%esi)
-        movl    %edx,           28(%esi)
-        /* Round 16 */
-        movl    16(%edi),       %eax
-        mull    %eax
-        movl    %eax,           32(%esi)
-        movl    %edx,           36(%esi)
-        /* Round 20 */
-        movl    20(%edi),       %eax
-        mull    %eax
-        movl    %eax,           40(%esi)
-        movl    %edx,           44(%esi)
-        /* Round 24 */
-        movl    24(%edi),       %eax
-        mull    %eax
-        movl    %eax,           48(%esi)
-        movl    %edx,           52(%esi)
-        /* Round 28 */
-        movl    28(%edi),       %eax
-        mull    %eax
-        movl    %eax,           56(%esi)
-        movl    %edx,           60(%esi)
-
-        addl    $32,            %edi
-        addl    $64,            %esi
-        subl    $8,             %ebx
-        jnz     .L009sw_loop
-.L008sw_finish:
-        movl    28(%esp),       %ebx
-        andl    $7,             %ebx
-        jz      .L010sw_end
-        /* Tail Round 0 */
-        movl    (%edi),         %eax
-        mull    %eax
-        movl    %eax,           (%esi)
-        decl    %ebx
-        movl    %edx,           4(%esi)
-        jz      .L010sw_end
-        /* Tail Round 1 */
-        movl    4(%edi),        %eax
-        mull    %eax
-        movl    %eax,           8(%esi)
-        decl    %ebx
-        movl    %edx,           12(%esi)
-        jz      .L010sw_end
-        /* Tail Round 2 */
-        movl    8(%edi),        %eax
-        mull    %eax
-        movl    %eax,           16(%esi)
-        decl    %ebx
-        movl    %edx,           20(%esi)
-        jz      .L010sw_end
-        /* Tail Round 3 */
-        movl    12(%edi),       %eax
-        mull    %eax
-        movl    %eax,           24(%esi)
-        decl    %ebx
-        movl    %edx,           28(%esi)
-        jz      .L010sw_end
-        /* Tail Round 4 */
-        movl    16(%edi),       %eax
-        mull    %eax
-        movl    %eax,           32(%esi)
-        decl    %ebx
-        movl    %edx,           36(%esi)
-        jz      .L010sw_end
-        /* Tail Round 5 */
-        movl    20(%edi),       %eax
-        mull    %eax
-        movl    %eax,           40(%esi)
-        decl    %ebx
-        movl    %edx,           44(%esi)
-        jz      .L010sw_end
-        /* Tail Round 6 */
-        movl    24(%edi),       %eax
-        mull    %eax
-        movl    %eax,           48(%esi)
-        movl    %edx,           52(%esi)
-.L010sw_end:
-        popl    %edi
-        popl    %esi
-        popl    %ebx
-        popl    %ebp
-        ret
-.bn_sqr_words_end:
-        SIZE(bn_sqr_words,.bn_sqr_words_end-bn_sqr_words)
-.ident  "bn_sqr_words"
-.text
-        .align ALIGN
-.globl bn_div64
-        TYPE(bn_div64,@function)
-bn_div64:
-        pushl   %ebp
-        pushl   %ebx
-        pushl   %esi
-        pushl   %edi
-
-        movl    20(%esp),       %edx
-        movl    24(%esp),       %eax
-        movl    28(%esp),       %ebx
-        divl    %ebx
-        popl    %edi
-        popl    %esi
-        popl    %ebx
-        popl    %ebp
-        ret
-.bn_div64_end:
-        SIZE(bn_div64,.bn_div64_end-bn_div64)
-.ident  "bn_div64"
-.text
-        .align ALIGN
-.globl bn_add_words
-        TYPE(bn_add_words,@function)
-bn_add_words:
-        pushl   %ebp
-        pushl   %ebx
-        pushl   %esi
-        pushl   %edi
-
-
-        movl    20(%esp),       %ebx
-        movl    24(%esp),       %esi
-        movl    28(%esp),       %edi
-        movl    32(%esp),       %ebp
-        xorl    %eax,           %eax
-        andl    $4294967288,    %ebp
-        jz      .L011aw_finish
-.L012aw_loop:
-        /* Round 0 */
-        movl    (%esi),         %ecx
-        movl    (%edi),         %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        movl    %ecx,           (%ebx)
-        /* Round 1 */
-        movl    4(%esi),        %ecx
-        movl    4(%edi),        %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        movl    %ecx,           4(%ebx)
-        /* Round 2 */
-        movl    8(%esi),        %ecx
-        movl    8(%edi),        %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        movl    %ecx,           8(%ebx)
-        /* Round 3 */
-        movl    12(%esi),       %ecx
-        movl    12(%edi),       %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        movl    %ecx,           12(%ebx)
-        /* Round 4 */
-        movl    16(%esi),       %ecx
-        movl    16(%edi),       %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        movl    %ecx,           16(%ebx)
-        /* Round 5 */
-        movl    20(%esi),       %ecx
-        movl    20(%edi),       %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        movl    %ecx,           20(%ebx)
-        /* Round 6 */
-        movl    24(%esi),       %ecx
-        movl    24(%edi),       %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        movl    %ecx,           24(%ebx)
-        /* Round 7 */
-        movl    28(%esi),       %ecx
-        movl    28(%edi),       %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        movl    %ecx,           28(%ebx)
-
-        addl    $32,            %esi
-        addl    $32,            %edi
-        addl    $32,            %ebx
-        subl    $8,             %ebp
-        jnz     .L012aw_loop
-.L011aw_finish:
-        movl    32(%esp),       %ebp
-        andl    $7,             %ebp
-        jz      .L013aw_end
-        /* Tail Round 0 */
-        movl    (%esi),         %ecx
-        movl    (%edi),         %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        decl    %ebp
-        movl    %ecx,           (%ebx)
-        jz      .L013aw_end
-        /* Tail Round 1 */
-        movl    4(%esi),        %ecx
-        movl    4(%edi),        %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        decl    %ebp
-        movl    %ecx,           4(%ebx)
-        jz      .L013aw_end
-        /* Tail Round 2 */
-        movl    8(%esi),        %ecx
-        movl    8(%edi),        %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        decl    %ebp
-        movl    %ecx,           8(%ebx)
-        jz      .L013aw_end
-        /* Tail Round 3 */
-        movl    12(%esi),       %ecx
-        movl    12(%edi),       %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        decl    %ebp
-        movl    %ecx,           12(%ebx)
-        jz      .L013aw_end
-        /* Tail Round 4 */
-        movl    16(%esi),       %ecx
-        movl    16(%edi),       %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        decl    %ebp
-        movl    %ecx,           16(%ebx)
-        jz      .L013aw_end
-        /* Tail Round 5 */
-        movl    20(%esi),       %ecx
-        movl    20(%edi),       %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        decl    %ebp
-        movl    %ecx,           20(%ebx)
-        jz      .L013aw_end
-        /* Tail Round 6 */
-        movl    24(%esi),       %ecx
-        movl    24(%edi),       %edx
-        addl    %eax,           %ecx
-        movl    $0,             %eax
-        adcl    %eax,           %eax
-        addl    %edx,           %ecx
-        adcl    $0,             %eax
-        movl    %ecx,           24(%ebx)
-.L013aw_end:
-        movl    %eax,           %eax
-        popl    %edi
-        popl    %esi
-        popl    %ebx
-        popl    %ebp
-        ret
-.bn_add_words_end:
-        SIZE(bn_add_words,.bn_add_words_end-bn_add_words)
-.ident  "bn_add_words"
diff --git a/src/lib/libcrypto/bn/asm/ca.pl b/src/lib/libcrypto/bn/asm/ca.pl
new file mode 100644
index 0000000000..c1ce67a6b4
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/ca.pl
@@ -0,0 +1,33 @@
+#!/usr/local/bin/perl
+# I have this in perl so I can use more usefull register names and then convert
+# them into alpha registers.
+#
+
+push(@INC,"perlasm","../../perlasm");
+require "alpha.pl";
+require "alpha/mul_add.pl";
+require "alpha/mul.pl";
+require "alpha/sqr.pl";
+require "alpha/add.pl";
+require "alpha/sub.pl";
+require "alpha/mul_c8.pl";
+require "alpha/mul_c4.pl";
+require "alpha/sqr_c4.pl";
+require "alpha/sqr_c8.pl";
+require "alpha/div.pl";
+
+&asm_init($ARGV[0],$0);
+
+&bn_mul_words("bn_mul_words");
+&bn_sqr_words("bn_sqr_words");
+&bn_mul_add_words("bn_mul_add_words");
+&bn_add_words("bn_add_words");
+&bn_sub_words("bn_sub_words");
+&bn_div_words("bn_div_words");
+&bn_mul_comba8("bn_mul_comba8");
+&bn_mul_comba4("bn_mul_comba4");
+&bn_sqr_comba4("bn_sqr_comba4");
+&bn_sqr_comba8("bn_sqr_comba8");
+
+&asm_finish();
+
diff --git a/src/lib/libcrypto/bn/asm/co-586.pl b/src/lib/libcrypto/bn/asm/co-586.pl
new file mode 100644
index 0000000000..5d962cb957
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/co-586.pl
@@ -0,0 +1,286 @@
+#!/usr/local/bin/perl
+
+push(@INC,"perlasm","../../perlasm");
+require "x86asm.pl";
+
+&asm_init($ARGV[0],$0);
+
+&bn_mul_comba("bn_mul_comba8",8);
+&bn_mul_comba("bn_mul_comba4",4);
+&bn_sqr_comba("bn_sqr_comba8",8);
+&bn_sqr_comba("bn_sqr_comba4",4);
+
+&asm_finish();
+
+sub mul_add_c
+        {
+        local($a,$ai,$b,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
+
+        # pos == -1 if eax and edx are pre-loaded, 0 to load from next
+        # words, and 1 if load return value
+
+        &comment("mul a[$ai]*b[$bi]");
+
+        # "eax" and "edx" will always be pre-loaded.
+        # &mov("eax",&DWP($ai*4,$a,"",0)) ;
+        # &mov("edx",&DWP($bi*4,$b,"",0));
+
+        &mul("edx");
+        &add($c0,"eax");
+         &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0;        # laod next a
+         &mov("eax",&wparam(0)) if $pos > 0;                    # load r[]
+         ###
+        &adc($c1,"edx");
+         &mov("edx",&DWP(($nb)*4,$b,"",0)) if $pos == 0;        # laod next b
+         &mov("edx",&DWP(($nb)*4,$b,"",0)) if $pos == 1;        # laod next b
+         ###
+        &adc($c2,0);
+         # is pos > 1, it means it is the last loop 
+         &mov(&DWP($i*4,"eax","",0),$c0) if $pos > 0;           # save r[];
+        &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1;         # laod next a
+        }
+
+sub sqr_add_c
+        {
+        local($r,$a,$ai,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
+
+        # pos == -1 if eax and edx are pre-loaded, 0 to load from next
+        # words, and 1 if load return value
+
+        &comment("sqr a[$ai]*a[$bi]");
+
+        # "eax" and "edx" will always be pre-loaded.
+        # &mov("eax",&DWP($ai*4,$a,"",0)) ;
+        # &mov("edx",&DWP($bi*4,$b,"",0));
+
+        if ($ai == $bi)
+                { &mul("eax");}
+        else
+                { &mul("edx");}
+        &add($c0,"eax");
+         &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0;        # load next a
+         ###
+        &adc($c1,"edx");
+         &mov("edx",&DWP(($nb)*4,$a,"",0)) if ($pos == 1) && ($na != $nb);
+         ###
+        &adc($c2,0);
+         # is pos > 1, it means it is the last loop 
+         &mov(&DWP($i*4,$r,"",0),$c0) if $pos > 0;              # save r[];
+        &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1;         # load next b
+        }
+
+sub sqr_add_c2
+        {
+        local($r,$a,$ai,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
+
+        # pos == -1 if eax and edx are pre-loaded, 0 to load from next
+        # words, and 1 if load return value
+
+        &comment("sqr a[$ai]*a[$bi]");
+
+        # "eax" and "edx" will always be pre-loaded.
+        # &mov("eax",&DWP($ai*4,$a,"",0)) ;
+        # &mov("edx",&DWP($bi*4,$a,"",0));
+
+        if ($ai == $bi)
+                { &mul("eax");}
+        else
+                { &mul("edx");}
+        &add("eax","eax");
+         ###
+        &adc("edx","edx");
+         ###
+        &adc($c2,0);
+         &add($c0,"eax");
+        &adc($c1,"edx");
+         &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0;        # load next a
+         &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1;        # load next b
+        &adc($c2,0);
+        &mov(&DWP($i*4,$r,"",0),$c0) if $pos > 0;               # save r[];
+         &mov("edx",&DWP(($nb)*4,$a,"",0)) if ($pos <= 1) && ($na != $nb);
+         ###
+        }
+
+sub bn_mul_comba
+        {
+        local($name,$num)=@_;
+        local($a,$b,$c0,$c1,$c2);
+        local($i,$as,$ae,$bs,$be,$ai,$bi);
+        local($tot,$end);
+
+        &function_begin_B($name,"");
+
+        $c0="ebx";
+        $c1="ecx";
+        $c2="ebp";
+        $a="esi";
+        $b="edi";
+        
+        $as=0;
+        $ae=0;
+        $bs=0;
+        $be=0;
+        $tot=$num+$num-1;
+
+        &push("esi");
+         &mov($a,&wparam(1));
+        &push("edi");
+         &mov($b,&wparam(2));
+        &push("ebp");
+         &push("ebx");
+
+        &xor($c0,$c0);
+         &mov("eax",&DWP(0,$a,"",0));   # load the first word 
+        &xor($c1,$c1);
+         &mov("edx",&DWP(0,$b,"",0));   # load the first second 
+
+        for ($i=0; $i<$tot; $i++)
+                {
+                $ai=$as;
+                $bi=$bs;
+                $end=$be+1;
+
+                &comment("################## Calculate word $i"); 
+
+                for ($j=$bs; $j<$end; $j++)
+                        {
+                        &xor($c2,$c2) if ($j == $bs);
+                        if (($j+1) == $end)
+                                {
+                                $v=1;
+                                $v=2 if (($i+1) == $tot);
+                                }
+                        else
+                                { $v=0; }
+                        if (($j+1) != $end)
+                                {
+                                $na=($ai-1);
+                                $nb=($bi+1);
+                                }
+                        else
+                                {
+                                $na=$as+($i < ($num-1));
+                                $nb=$bs+($i >= ($num-1));
+                                }
+#printf STDERR "[$ai,$bi] -> [$na,$nb]\n";
+                        &mul_add_c($a,$ai,$b,$bi,$c0,$c1,$c2,$v,$i,$na,$nb);
+                        if ($v)
+                                {
+                                &comment("saved r[$i]");
+                                # &mov("eax",&wparam(0));
+                                # &mov(&DWP($i*4,"eax","",0),$c0);
+                                ($c0,$c1,$c2)=($c1,$c2,$c0);
+                                }
+                        $ai--;
+                        $bi++;
+                        }
+                $as++ if ($i < ($num-1));
+                $ae++ if ($i >= ($num-1));
+
+                $bs++ if ($i >= ($num-1));
+                $be++ if ($i < ($num-1));
+                }
+        &comment("save r[$i]");
+        # &mov("eax",&wparam(0));
+        &mov(&DWP($i*4,"eax","",0),$c0);
+
+        &pop("ebx");
+        &pop("ebp");
+        &pop("edi");
+        &pop("esi");
+        &ret();
+        &function_end_B($name);
+        }
+
+sub bn_sqr_comba
+        {
+        local($name,$num)=@_;
+        local($r,$a,$c0,$c1,$c2)=@_;
+        local($i,$as,$ae,$bs,$be,$ai,$bi);
+        local($b,$tot,$end,$half);
+
+        &function_begin_B($name,"");
+
+        $c0="ebx";
+        $c1="ecx";
+        $c2="ebp";
+        $a="esi";
+        $r="edi";
+
+        &push("esi");
+         &push("edi");
+        &push("ebp");
+         &push("ebx");
+        &mov($r,&wparam(0));
+         &mov($a,&wparam(1));
+        &xor($c0,$c0);
+         &xor($c1,$c1);
+        &mov("eax",&DWP(0,$a,"",0)); # load the first word
+
+        $as=0;
+        $ae=0;
+        $bs=0;
+        $be=0;
+        $tot=$num+$num-1;
+
+        for ($i=0; $i<$tot; $i++)
+                {
+                $ai=$as;
+                $bi=$bs;
+                $end=$be+1;
+
+                &comment("############### Calculate word $i");
+                for ($j=$bs; $j<$end; $j++)
+                        {
+                        &xor($c2,$c2) if ($j == $bs);
+                        if (($ai-1) < ($bi+1))
+                                {
+                                $v=1;
+                                $v=2 if ($i+1) == $tot;
+                                }
+                        else
+                                { $v=0; }
+                        if (!$v)
+                                {
+                                $na=$ai-1;
+                                $nb=$bi+1;
+                                }
+                        else
+                                {
+                                $na=$as+($i < ($num-1));
+                                $nb=$bs+($i >= ($num-1));
+                                }
+                        if ($ai == $bi)
+                                {
+                                &sqr_add_c($r,$a,$ai,$bi,
+                                        $c0,$c1,$c2,$v,$i,$na,$nb);
+                                }
+                        else
+                                {
+                                &sqr_add_c2($r,$a,$ai,$bi,
+                                        $c0,$c1,$c2,$v,$i,$na,$nb);
+                                }
+                        if ($v)
+                                {
+                                &comment("saved r[$i]");
+                                #&mov(&DWP($i*4,$r,"",0),$c0);
+                                ($c0,$c1,$c2)=($c1,$c2,$c0);
+                                last;
+                                }
+                        $ai--;
+                        $bi++;
+                        }
+                $as++ if ($i < ($num-1));
+                $ae++ if ($i >= ($num-1));
+
+                $bs++ if ($i >= ($num-1));
+                $be++ if ($i < ($num-1));
+                }
+        &mov(&DWP($i*4,$r,"",0),$c0);
+        &pop("ebx");
+        &pop("ebp");
+        &pop("edi");
+        &pop("esi");
+        &ret();
+        &function_end_B($name);
+        }
diff --git a/src/lib/libcrypto/bn/asm/co-alpha.pl b/src/lib/libcrypto/bn/asm/co-alpha.pl
new file mode 100644
index 0000000000..67dad3e3d5
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/co-alpha.pl
@@ -0,0 +1,116 @@
+#!/usr/local/bin/perl
+# I have this in perl so I can use more usefull register names and then convert
+# them into alpha registers.
+#
+
+push(@INC,"perlasm","../../perlasm");
+require "alpha.pl";
+
+&asm_init($ARGV[0],$0);
+
+print &bn_sub_words("bn_sub_words");
+
+&asm_finish();
+
+sub bn_sub_words
+        {
+        local($name)=@_;
+        local($cc,$a,$b,$r);
+
+        $cc="r0";
+        $a0="r1"; $b0="r5"; $r0="r9";  $tmp="r13";
+        $a1="r2"; $b1="r6"; $r1="r10"; $t1="r14";
+        $a2="r3"; $b2="r7"; $r2="r11";
+        $a3="r4"; $b3="r8"; $r3="r12"; $t3="r15";
+
+        $rp=&wparam(0);
+        $ap=&wparam(1);
+        $bp=&wparam(2);
+        $count=&wparam(3);
+
+        &function_begin($name,"");
+
+        &comment("");
+        &sub($count,4,$count);
+        &mov("zero",$cc);
+        &blt($count,&label("finish"));
+
+        &ld($a0,&QWPw(0,$ap));
+        &ld($b0,&QWPw(0,$bp));
+
+##########################################################
+        &set_label("loop");
+
+        &ld($a1,&QWPw(1,$ap));
+         &cmpult($a0,$b0,$tmp); # will we borrow?
+        &ld($b1,&QWPw(1,$bp));
+         &sub($a0,$b0,$a0);             # do the subtract
+        &ld($a2,&QWPw(2,$ap));
+         &cmpult($a0,$cc,$b0);  # will we borrow?
+        &ld($b2,&QWPw(2,$bp));
+         &sub($a0,$cc,$a0);     # will we borrow?
+        &ld($a3,&QWPw(3,$ap));
+         &add($b0,$tmp,$cc);    # add the borrows
+
+        &cmpult($a1,$b1,$t1);   # will we borrow?
+         &sub($a1,$b1,$a1);     # do the subtract
+        &ld($b3,&QWPw(3,$bp));
+         &cmpult($a1,$cc,$b1);  # will we borrow?
+        &sub($a1,$cc,$a1);      # will we borrow?
+         &add($b1,$t1,$cc);     # add the borrows
+
+        &cmpult($a2,$b2,$tmp);  # will we borrow?
+         &sub($a2,$b2,$a2);             # do the subtract
+        &st($a0,&QWPw(0,$rp));  # save
+         &cmpult($a2,$cc,$b2);  # will we borrow?
+        &sub($a2,$cc,$a2);      # will we borrow?
+         &add($b2,$tmp,$cc);    # add the borrows
+
+        &cmpult($a3,$b3,$t3);   # will we borrow?
+         &sub($a3,$b3,$a3);             # do the subtract
+        &st($a1,&QWPw(1,$rp));  # save
+         &cmpult($a3,$cc,$b3);  # will we borrow?
+        &sub($a3,$cc,$a3);      # will we borrow?
+         &add($b3,$t3,$cc);     # add the borrows
+
+        &st($a2,&QWPw(2,$rp));  # save
+         &sub($count,4,$count); # count-=4
+        &st($a3,&QWPw(3,$rp));  # save
+         &add($ap,4*$QWS,$ap);  # count+=4
+        &add($bp,4*$QWS,$bp);   # count+=4
+         &add($rp,4*$QWS,$rp);  # count+=4
+
+        &blt($count,&label("finish"));
+        &ld($a0,&QWPw(0,$ap));
+         &ld($b0,&QWPw(0,$bp));
+        &br(&label("loop"));
+##################################################
+        # Do the last 0..3 words
+
+        &set_label("last_loop");
+
+        &ld($a0,&QWPw(0,$ap));  # get a
+         &ld($b0,&QWPw(0,$bp)); # get b
+        &cmpult($a0,$b0,$tmp);  # will we borrow?
+        &sub($a0,$b0,$a0);      # do the subtract
+        &cmpult($a0,$cc,$b0);   # will we borrow?
+        &sub($a0,$cc,$a0);      # will we borrow?
+        &st($a0,&QWPw(0,$rp));  # save
+        &add($b0,$tmp,$cc);     # add the borrows
+
+        &add($ap,$QWS,$ap);
+        &add($bp,$QWS,$bp);
+        &add($rp,$QWS,$rp);
+        &sub($count,1,$count);
+        &bgt($count,&label("last_loop"));
+        &function_end_A($name);
+
+######################################################
+        &set_label("finish");
+        &add($count,4,$count);
+        &bgt($count,&label("last_loop"));
+
+        &set_label("end");
+        &function_end($name);
+        }
+
diff --git a/src/lib/libcrypto/bn/asm/mips1.s b/src/lib/libcrypto/bn/asm/mips1.s
new file mode 100644
index 0000000000..44fa1254c7
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/mips1.s
@@ -0,0 +1,539 @@
+/* This assember is for R2000/R3000 machines, or higher ones that do
+ * no want to do any 64 bit arithmatic.
+ * Make sure that the SSLeay bignum library is compiled with 
+ * THIRTY_TWO_BIT set.
+ * This must either be compiled with the system CC, or, if you use GNU gas,
+ * cc -E mips1.s|gas -o mips1.o
+ */
+        .set    reorder
+        .set    noat
+
+#define R1      $1
+#define CC      $2
+#define R2      $3
+#define R3      $8
+#define R4      $9
+#define L1      $10
+#define L2      $11
+#define L3      $12
+#define L4      $13
+#define H1      $14
+#define H2      $15
+#define H3      $24
+#define H4      $25
+
+#define P1      $4
+#define P2      $5
+#define P3      $6
+#define P4      $7
+
+        .align  2
+        .ent    bn_mul_add_words
+        .globl  bn_mul_add_words
+.text
+bn_mul_add_words:
+        .frame  $sp,0,$31
+        .mask   0x00000000,0
+        .fmask  0x00000000,0
+
+        #blt    P3,4,$lab34
+        
+        subu    R1,P3,4
+        move    CC,$0
+        bltz    R1,$lab34
+$lab2:  
+        lw      R1,0(P1)
+         lw     L1,0(P2)
+        lw      R2,4(P1)
+         lw     L2,4(P2)
+        lw      R3,8(P1)
+         lw     L3,8(P2)
+        lw      R4,12(P1)
+         lw     L4,12(P2)
+        multu   L1,P4
+         addu   R1,R1,CC
+        mflo    L1
+         sltu   CC,R1,CC
+        addu    R1,R1,L1
+         mfhi   H1
+        sltu    L1,R1,L1
+         sw     R1,0(P1)
+        addu    CC,CC,L1
+         multu  L2,P4
+        addu    CC,H1,CC
+        mflo    L2
+         addu   R2,R2,CC
+        sltu    CC,R2,CC
+         mfhi   H2
+        addu    R2,R2,L2
+         addu   P2,P2,16
+        sltu    L2,R2,L2
+         sw     R2,4(P1)
+        addu    CC,CC,L2
+         multu  L3,P4
+        addu    CC,H2,CC
+        mflo    L3
+         addu   R3,R3,CC
+        sltu    CC,R3,CC
+         mfhi   H3
+        addu    R3,R3,L3
+         addu   P1,P1,16
+        sltu    L3,R3,L3
+         sw     R3,-8(P1)
+        addu    CC,CC,L3
+         multu  L4,P4
+        addu    CC,H3,CC
+        mflo    L4
+         addu   R4,R4,CC
+        sltu    CC,R4,CC
+         mfhi   H4
+        addu    R4,R4,L4
+         subu   P3,P3,4
+        sltu    L4,R4,L4
+        addu    CC,CC,L4
+        addu    CC,H4,CC
+
+        subu    R1,P3,4
+        sw      R4,-4(P1)       # delay slot
+        bgez    R1,$lab2
+
+        bleu    P3,0,$lab3
+        .align  2
+$lab33: 
+        lw      L1,0(P2)
+         lw     R1,0(P1)
+        multu   L1,P4
+         addu   R1,R1,CC
+        sltu    CC,R1,CC
+         addu   P1,P1,4
+        mflo    L1
+         mfhi   H1
+        addu    R1,R1,L1
+         addu   P2,P2,4
+        sltu    L1,R1,L1
+         subu   P3,P3,1
+        addu    CC,CC,L1
+         sw     R1,-4(P1)
+        addu    CC,H1,CC
+         bgtz   P3,$lab33
+        j       $31
+        .align  2
+$lab3:
+        j       $31
+        .align  2
+$lab34:
+        bgt     P3,0,$lab33
+        j       $31
+        .end    bn_mul_add_words
+
+        .align  2
+        # Program Unit: bn_mul_words
+        .ent    bn_mul_words
+        .globl  bn_mul_words
+.text
+bn_mul_words:
+        .frame  $sp,0,$31
+        .mask   0x00000000,0
+        .fmask  0x00000000,0
+        
+        subu    P3,P3,4
+        move    CC,$0
+        bltz    P3,$lab45
+$lab44: 
+        lw      L1,0(P2)
+         lw     L2,4(P2)
+        lw      L3,8(P2)
+         lw     L4,12(P2)
+        multu   L1,P4
+         subu   P3,P3,4
+        mflo    L1
+         mfhi   H1
+        addu    L1,L1,CC
+         multu  L2,P4
+        sltu    CC,L1,CC
+         sw     L1,0(P1)
+        addu    CC,H1,CC
+         mflo   L2
+        mfhi    H2
+         addu   L2,L2,CC
+        multu   L3,P4
+         sltu   CC,L2,CC
+        sw      L2,4(P1)
+         addu   CC,H2,CC
+        mflo    L3
+         mfhi   H3
+        addu    L3,L3,CC
+         multu  L4,P4
+        sltu    CC,L3,CC
+         sw     L3,8(P1)
+        addu    CC,H3,CC
+         mflo   L4
+        mfhi    H4
+         addu   L4,L4,CC
+        addu    P1,P1,16
+         sltu   CC,L4,CC
+        addu    P2,P2,16
+         addu   CC,H4,CC
+        sw      L4,-4(P1)
+
+        bgez    P3,$lab44
+        b       $lab45
+$lab46:
+        lw      L1,0(P2)
+         addu   P1,P1,4
+        multu   L1,P4
+         addu   P2,P2,4
+        mflo    L1
+         mfhi   H1
+        addu    L1,L1,CC
+         subu   P3,P3,1
+        sltu    CC,L1,CC
+         sw     L1,-4(P1)
+        addu    CC,H1,CC
+         bgtz   P3,$lab46
+        j       $31
+$lab45:
+        addu    P3,P3,4
+        bgtz    P3,$lab46
+        j       $31
+        .align  2
+        .end    bn_mul_words
+
+        # Program Unit: bn_sqr_words
+        .ent    bn_sqr_words
+        .globl  bn_sqr_words
+.text
+bn_sqr_words:
+        .frame  $sp,0,$31
+        .mask   0x00000000,0
+        .fmask  0x00000000,0
+        
+        subu    P3,P3,4
+        bltz    P3,$lab55
+$lab54:
+        lw      L1,0(P2)
+         lw     L2,4(P2)
+        lw      L3,8(P2)
+         lw     L4,12(P2)
+
+        multu   L1,L1
+         subu   P3,P3,4
+        mflo    L1
+         mfhi   H1
+        sw      L1,0(P1)
+         sw     H1,4(P1)
+
+        multu   L2,L2
+         addu   P1,P1,32
+        mflo    L2
+         mfhi   H2
+        sw      L2,-24(P1)
+         sw     H2,-20(P1)
+
+        multu   L3,L3
+         addu   P2,P2,16
+        mflo    L3
+         mfhi   H3
+        sw      L3,-16(P1)
+         sw     H3,-12(P1)
+
+        multu   L4,L4
+
+        mflo    L4
+         mfhi   H4
+        sw      L4,-8(P1)
+         sw     H4,-4(P1)
+
+        bgtz    P3,$lab54
+        b       $lab55
+$lab56: 
+        lw      L1,0(P2)
+        addu    P1,P1,8
+        multu   L1,L1
+        addu    P2,P2,4
+        subu    P3,P3,1
+        mflo    L1
+        mfhi    H1
+        sw      L1,-8(P1)
+        sw      H1,-4(P1)
+
+        bgtz    P3,$lab56
+        j       $31
+$lab55:
+        addu    P3,P3,4
+        bgtz    P3,$lab56
+        j       $31
+        .align  2
+        .end    bn_sqr_words
+
+        # Program Unit: bn_add_words
+        .ent    bn_add_words
+        .globl  bn_add_words
+.text
+bn_add_words:    # 0x590
+        .frame  $sp,0,$31
+        .mask   0x00000000,0
+        .fmask  0x00000000,0
+        
+        subu    P4,P4,4
+        move    CC,$0
+        bltz    P4,$lab65
+$lab64: 
+        lw      L1,0(P2)
+        lw      R1,0(P3)
+        lw      L2,4(P2)
+        lw      R2,4(P3)
+
+        addu    L1,L1,CC
+         lw     L3,8(P2)
+        sltu    CC,L1,CC
+         addu   L1,L1,R1
+        sltu    R1,L1,R1
+         lw     R3,8(P3)
+        addu    CC,CC,R1
+         lw     L4,12(P2)
+
+        addu    L2,L2,CC
+         lw     R4,12(P3)
+        sltu    CC,L2,CC
+         addu   L2,L2,R2
+        sltu    R2,L2,R2
+         sw     L1,0(P1)
+        addu    CC,CC,R2
+         addu   P1,P1,16
+        addu    L3,L3,CC
+         sw     L2,-12(P1)
+ 
+        sltu    CC,L3,CC
+         addu   L3,L3,R3
+        sltu    R3,L3,R3
+         addu   P2,P2,16
+        addu    CC,CC,R3
+
+        addu    L4,L4,CC
+         addu   P3,P3,16
+        sltu    CC,L4,CC
+         addu   L4,L4,R4
+        subu    P4,P4,4
+         sltu   R4,L4,R4
+        sw      L3,-8(P1)
+         addu   CC,CC,R4
+        sw      L4,-4(P1)
+
+        bgtz    P4,$lab64
+        b       $lab65
+$lab66:
+        lw      L1,0(P2)
+         lw     R1,0(P3)
+        addu    L1,L1,CC
+         addu   P1,P1,4
+        sltu    CC,L1,CC
+         addu   P2,P2,4
+        addu    P3,P3,4
+         addu   L1,L1,R1
+        subu    P4,P4,1
+         sltu   R1,L1,R1
+        sw      L1,-4(P1)
+         addu   CC,CC,R1
+
+        bgtz    P4,$lab66
+        j       $31
+$lab65:
+        addu    P4,P4,4
+        bgtz    P4,$lab66
+        j       $31
+        .end    bn_add_words
+
+        # Program Unit: bn_div64
+        .set    at
+        .set    reorder
+        .text   
+        .align  2
+        .globl  bn_div64
+ # 321          {
+        .ent    bn_div64 2
+bn_div64:
+        subu    $sp, 64
+        sw      $31, 56($sp)
+        sw      $16, 48($sp)
+        .mask   0x80010000, -56
+        .frame  $sp, 64, $31
+        move    $9, $4
+        move    $12, $5
+        move    $16, $6
+ # 322          BN_ULONG dh,dl,q,ret=0,th,tl,t;
+        move    $31, $0
+ # 323          int i,count=2;
+        li      $13, 2
+ # 324  
+ # 325          if (d == 0) return(BN_MASK2);
+        bne     $16, 0, $80
+        li      $2, -1
+        b       $93
+$80:
+ # 326  
+ # 327          i=BN_num_bits_word(d);
+        move    $4, $16
+        sw      $31, 16($sp)
+        sw      $9, 24($sp)
+        sw      $12, 32($sp)
+        sw      $13, 40($sp)
+        .livereg        0x800ff0e,0xfff
+        jal     BN_num_bits_word
+        li      $4, 32
+        lw      $31, 16($sp)
+        lw      $9, 24($sp)
+        lw      $12, 32($sp)
+        lw      $13, 40($sp)
+        move    $3, $2
+ # 328          if ((i != BN_BITS2) && (h > (BN_ULONG)1<<i))
+        beq     $2, $4, $81
+        li      $14, 1
+        sll     $15, $14, $2
+        bleu    $9, $15, $81
+ # 329                  {
+ # 330  #if !defined(NO_STDIO) && !defined(WIN16)
+ # 331                  fprintf(stderr,"Division would overflow (%d)\n",i);
+ # 332  #endif
+ # 333                  abort();
+        sw      $3, 8($sp)
+        sw      $9, 24($sp)
+        sw      $12, 32($sp)
+        sw      $13, 40($sp)
+        sw      $31, 26($sp)
+        .livereg        0xff0e,0xfff
+        jal     abort
+        lw      $3, 8($sp)
+        li      $4, 32
+        lw      $9, 24($sp)
+        lw      $12, 32($sp)
+        lw      $13, 40($sp)
+        lw      $31, 26($sp)
+ # 334                  }
+$81:
+ # 335          i=BN_BITS2-i;
+        subu    $3, $4, $3
+ # 336          if (h >= d) h-=d;
+        bltu    $9, $16, $82
+        subu    $9, $9, $16
+$82:
+ # 337  
+ # 338          if (i)
+        beq     $3, 0, $83
+ # 339                  {
+ # 340                  d<<=i;
+        sll     $16, $16, $3
+ # 341                  h=(h<<i)|(l>>(BN_BITS2-i));
+        sll     $24, $9, $3
+        subu    $25, $4, $3
+        srl     $14, $12, $25
+        or      $9, $24, $14
+ # 342                  l<<=i;
+        sll     $12, $12, $3
+ # 343                  }
+$83:
+ # 344          dh=(d&BN_MASK2h)>>BN_BITS4;
+ # 345          dl=(d&BN_MASK2l);
+        and     $8, $16, -65536
+        srl     $8, $8, 16
+        and     $10, $16, 65535
+        li      $6, -65536
+$84:
+ # 346          for (;;)
+ # 347                  {
+ # 348                  if ((h>>BN_BITS4) == dh)
+        srl     $15, $9, 16
+        bne     $8, $15, $85
+ # 349                          q=BN_MASK2l;
+        li      $5, 65535
+        b       $86
+$85:
+ # 350                  else
+ # 351                          q=h/dh;
+        divu    $5, $9, $8
+$86:
+ # 352  
+ # 353                  for (;;)
+ # 354                          {
+ # 355                          t=(h-q*dh);
+        mul     $4, $5, $8
+        subu    $2, $9, $4
+        move    $3, $2
+ # 356                          if ((t&BN_MASK2h) ||
+ # 357                                  ((dl*q) <= (
+ # 358                                          (t<<BN_BITS4)+
+ # 359                                          ((l&BN_MASK2h)>>BN_BITS4))))
+        and     $25, $2, $6
+        bne     $25, $0, $87
+        mul     $24, $10, $5
+        sll     $14, $3, 16
+        and     $15, $12, $6
+        srl     $25, $15, 16
+        addu    $15, $14, $25
+        bgtu    $24, $15, $88
+$87:
+ # 360                                  break;
+        mul     $3, $10, $5
+        b       $89
+$88:
+ # 361                          q--;
+        addu    $5, $5, -1
+ # 362                          }
+        b       $86
+$89:
+ # 363                  th=q*dh;
+ # 364                  tl=q*dl;
+ # 365                  t=(tl>>BN_BITS4);
+ # 366                  tl=(tl<<BN_BITS4)&BN_MASK2h;
+        sll     $14, $3, 16
+        and     $2, $14, $6
+        move    $11, $2
+ # 367                  th+=t;
+        srl     $25, $3, 16
+        addu    $7, $4, $25
+ # 368  
+ # 369                  if (l < tl) th++;
+        bgeu    $12, $2, $90
+        addu    $7, $7, 1
+$90:
+ # 370                  l-=tl;
+        subu    $12, $12, $11
+ # 371                  if (h < th)
+        bgeu    $9, $7, $91
+ # 372                          {
+ # 373                          h+=d;
+        addu    $9, $9, $16
+ # 374                          q--;
+        addu    $5, $5, -1
+ # 375                          }
+$91:
+ # 376                  h-=th;
+        subu    $9, $9, $7
+ # 377  
+ # 378                  if (--count == 0) break;
+        addu    $13, $13, -1
+        beq     $13, 0, $92
+ # 379  
+ # 380                  ret=q<<BN_BITS4;
+        sll     $31, $5, 16
+ # 381                  h=((h<<BN_BITS4)|(l>>BN_BITS4))&BN_MASK2;
+        sll     $24, $9, 16
+        srl     $15, $12, 16
+        or      $9, $24, $15
+ # 382                  l=(l&BN_MASK2l)<<BN_BITS4;
+        and     $12, $12, 65535
+        sll     $12, $12, 16
+ # 383                  }
+        b       $84
+$92:
+ # 384          ret|=q;
+        or      $31, $31, $5
+ # 385          return(ret);
+        move    $2, $31
+$93:
+        lw      $16, 48($sp)
+        lw      $31, 56($sp)
+        addu    $sp, 64
+        j       $31
+        .end    bn_div64
+
diff --git a/src/lib/libcrypto/bn/asm/mips3.s b/src/lib/libcrypto/bn/asm/mips3.s
new file mode 100644
index 0000000000..191345d920
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/mips3.s
@@ -0,0 +1,2138 @@
+.rdata
+.asciiz "mips3.s, Version 1.0"
+.asciiz "MIPS III/IV ISA artwork by Andy Polyakov <appro@fy.chalmers.se>"
+
+/*
+ * ====================================================================
+ * Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+ * project.
+ *
+ * Rights for redistribution and usage in source and binary forms are
+ * granted according to the OpenSSL license. Warranty of any kind is
+ * disclaimed.
+ * ====================================================================
+ */
+
+/*
+ * This is my modest contributon to the OpenSSL project (see
+ * http://www.openssl.org/ for more information about it) and is
+ * a drop-in MIPS III/IV ISA replacement for crypto/bn/bn_asm.c
+ * module. For updates see http://fy.chalmers.se/~appro/hpe/.
+ *
+ * The module is designed to work with either of the "new" MIPS ABI(5),
+ * namely N32 or N64, offered by IRIX 6.x. It's not ment to work under
+ * IRIX 5.x not only because it doesn't support new ABIs but also
+ * because 5.x kernels put R4x00 CPU into 32-bit mode and all those
+ * 64-bit instructions (daddu, dmultu, etc.) found below gonna only
+ * cause illegal instruction exception:-(
+ *
+ * In addition the code depends on preprocessor flags set up by MIPSpro
+ * compiler driver (either as or cc) and therefore (probably?) can't be
+ * compiled by the GNU assembler. GNU C driver manages fine though...
+ * I mean as long as -mmips-as is specified or is the default option,
+ * because then it simply invokes /usr/bin/as which in turn takes
+ * perfect care of the preprocessor definitions. Another neat feature
+ * offered by the MIPSpro assembler is an optimization pass. This gave
+ * me the opportunity to have the code looking more regular as all those
+ * architecture dependent instruction rescheduling details were left to
+ * the assembler. Cool, huh?
+ *
+ * Performance improvement is astonishing! 'apps/openssl speed rsa dsa'
+ * goes way over 3 times faster!
+ *
+ *                                      <appro@fy.chalmers.se>
+ */
+#include <asm.h>
+#include <regdef.h>
+
+#if _MIPS_ISA>=4
+#define MOVNZ(cond,dst,src)     \
+        movn    dst,src,cond
+#else
+#define MOVNZ(cond,dst,src)     \
+        .set    noreorder;      \
+        bnezl   cond,.+8;       \
+        move    dst,src;        \
+        .set    reorder
+#endif
+
+.text
+
+.set    noat
+.set    reorder
+
+#define MINUS4  v1
+
+.align  5
+LEAF(bn_mul_add_words)
+        .set    noreorder
+        bgtzl   a2,.L_bn_mul_add_words_proceed
+        ld      t0,0(a1)
+        jr      ra
+        move    v0,zero
+        .set    reorder
+
+.L_bn_mul_add_words_proceed:
+        li      MINUS4,-4
+        and     ta0,a2,MINUS4
+        move    v0,zero
+        beqz    ta0,.L_bn_mul_add_words_tail
+
+.L_bn_mul_add_words_loop:
+        dmultu  t0,a3
+        ld      t1,0(a0)
+        ld      t2,8(a1)
+        ld      t3,8(a0)
+        ld      ta0,16(a1)
+        ld      ta1,16(a0)
+        daddu   t1,v0
+        sltu    v0,t1,v0        /* All manuals say it "compares 32-bit
+                                 * values", but it seems to work fine
+                                 * even on 64-bit registers. */
+        mflo    AT
+        mfhi    t0
+        daddu   t1,AT
+        daddu   v0,t0
+        sltu    AT,t1,AT
+        sd      t1,0(a0)
+        daddu   v0,AT
+
+        dmultu  t2,a3
+        ld      ta2,24(a1)
+        ld      ta3,24(a0)
+        daddu   t3,v0
+        sltu    v0,t3,v0
+        mflo    AT
+        mfhi    t2
+        daddu   t3,AT
+        daddu   v0,t2
+        sltu    AT,t3,AT
+        sd      t3,8(a0)
+        daddu   v0,AT
+
+        dmultu  ta0,a3
+        subu    a2,4
+        PTR_ADD a0,32
+        PTR_ADD a1,32
+        daddu   ta1,v0
+        sltu    v0,ta1,v0
+        mflo    AT
+        mfhi    ta0
+        daddu   ta1,AT
+        daddu   v0,ta0
+        sltu    AT,ta1,AT
+        sd      ta1,-16(a0)
+        daddu   v0,AT
+
+
+        dmultu  ta2,a3
+        and     ta0,a2,MINUS4
+        daddu   ta3,v0
+        sltu    v0,ta3,v0
+        mflo    AT
+        mfhi    ta2
+        daddu   ta3,AT
+        daddu   v0,ta2
+        sltu    AT,ta3,AT
+        sd      ta3,-8(a0)
+        daddu   v0,AT
+        .set    noreorder
+        bgtzl   ta0,.L_bn_mul_add_words_loop
+        ld      t0,0(a1)
+
+        bnezl   a2,.L_bn_mul_add_words_tail
+        ld      t0,0(a1)
+        .set    reorder
+
+.L_bn_mul_add_words_return:
+        jr      ra
+
+.L_bn_mul_add_words_tail:
+        dmultu  t0,a3
+        ld      t1,0(a0)
+        subu    a2,1
+        daddu   t1,v0
+        sltu    v0,t1,v0
+        mflo    AT
+        mfhi    t0
+        daddu   t1,AT
+        daddu   v0,t0
+        sltu    AT,t1,AT
+        sd      t1,0(a0)
+        daddu   v0,AT
+        beqz    a2,.L_bn_mul_add_words_return
+
+        ld      t0,8(a1)
+        dmultu  t0,a3
+        ld      t1,8(a0)
+        subu    a2,1
+        daddu   t1,v0
+        sltu    v0,t1,v0
+        mflo    AT
+        mfhi    t0
+        daddu   t1,AT
+        daddu   v0,t0
+        sltu    AT,t1,AT
+        sd      t1,8(a0)
+        daddu   v0,AT
+        beqz    a2,.L_bn_mul_add_words_return
+
+        ld      t0,16(a1)
+        dmultu  t0,a3
+        ld      t1,16(a0)
+        daddu   t1,v0
+        sltu    v0,t1,v0
+        mflo    AT
+        mfhi    t0
+        daddu   t1,AT
+        daddu   v0,t0
+        sltu    AT,t1,AT
+        sd      t1,16(a0)
+        daddu   v0,AT
+        jr      ra
+END(bn_mul_add_words)
+
+.align  5
+LEAF(bn_mul_words)
+        .set    noreorder
+        bgtzl   a2,.L_bn_mul_words_proceed
+        ld      t0,0(a1)
+        jr      ra
+        move    v0,zero
+        .set    reorder
+
+.L_bn_mul_words_proceed:
+        li      MINUS4,-4
+        and     ta0,a2,MINUS4
+        move    v0,zero
+        beqz    ta0,.L_bn_mul_words_tail
+
+.L_bn_mul_words_loop:
+        dmultu  t0,a3
+        ld      t2,8(a1)
+        ld      ta0,16(a1)
+        ld      ta2,24(a1)
+        mflo    AT
+        mfhi    t0
+        daddu   v0,AT
+        sltu    t1,v0,AT
+        sd      v0,0(a0)
+        daddu   v0,t1,t0
+
+        dmultu  t2,a3
+        subu    a2,4
+        PTR_ADD a0,32
+        PTR_ADD a1,32
+        mflo    AT
+        mfhi    t2
+        daddu   v0,AT
+        sltu    t3,v0,AT
+        sd      v0,-24(a0)
+        daddu   v0,t3,t2
+
+        dmultu  ta0,a3
+        mflo    AT
+        mfhi    ta0
+        daddu   v0,AT
+        sltu    ta1,v0,AT
+        sd      v0,-16(a0)
+        daddu   v0,ta1,ta0
+
+
+        dmultu  ta2,a3
+        and     ta0,a2,MINUS4
+        mflo    AT
+        mfhi    ta2
+        daddu   v0,AT
+        sltu    ta3,v0,AT
+        sd      v0,-8(a0)
+        daddu   v0,ta3,ta2
+        .set    noreorder
+        bgtzl   ta0,.L_bn_mul_words_loop
+        ld      t0,0(a1)
+
+        bnezl   a2,.L_bn_mul_words_tail
+        ld      t0,0(a1)
+        .set    reorder
+
+.L_bn_mul_words_return:
+        jr      ra
+
+.L_bn_mul_words_tail:
+        dmultu  t0,a3
+        subu    a2,1
+        mflo    AT
+        mfhi    t0
+        daddu   v0,AT
+        sltu    t1,v0,AT
+        sd      v0,0(a0)
+        daddu   v0,t1,t0
+        beqz    a2,.L_bn_mul_words_return
+
+        ld      t0,8(a1)
+        dmultu  t0,a3
+        subu    a2,1
+        mflo    AT
+        mfhi    t0
+        daddu   v0,AT
+        sltu    t1,v0,AT
+        sd      v0,8(a0)
+        daddu   v0,t1,t0
+        beqz    a2,.L_bn_mul_words_return
+
+        ld      t0,16(a1)
+        dmultu  t0,a3
+        mflo    AT
+        mfhi    t0
+        daddu   v0,AT
+        sltu    t1,v0,AT
+        sd      v0,16(a0)
+        daddu   v0,t1,t0
+        jr      ra
+END(bn_mul_words)
+
+.align  5
+LEAF(bn_sqr_words)
+        .set    noreorder
+        bgtzl   a2,.L_bn_sqr_words_proceed
+        ld      t0,0(a1)
+        jr      ra
+        move    v0,zero
+        .set    reorder
+
+.L_bn_sqr_words_proceed:
+        li      MINUS4,-4
+        and     ta0,a2,MINUS4
+        move    v0,zero
+        beqz    ta0,.L_bn_sqr_words_tail
+
+.L_bn_sqr_words_loop:
+        dmultu  t0,t0
+        ld      t2,8(a1)
+        ld      ta0,16(a1)
+        ld      ta2,24(a1)
+        mflo    t1
+        mfhi    t0
+        sd      t1,0(a0)
+        sd      t0,8(a0)
+
+        dmultu  t2,t2
+        subu    a2,4
+        PTR_ADD a0,64
+        PTR_ADD a1,32
+        mflo    t3
+        mfhi    t2
+        sd      t3,-48(a0)
+        sd      t2,-40(a0)
+
+        dmultu  ta0,ta0
+        mflo    ta1
+        mfhi    ta0
+        sd      ta1,-32(a0)
+        sd      ta0,-24(a0)
+
+
+        dmultu  ta2,ta2
+        and     ta0,a2,MINUS4
+        mflo    ta3
+        mfhi    ta2
+        sd      ta3,-16(a0)
+        sd      ta2,-8(a0)
+
+        .set    noreorder
+        bgtzl   ta0,.L_bn_sqr_words_loop
+        ld      t0,0(a1)
+
+        bnezl   a2,.L_bn_sqr_words_tail
+        ld      t0,0(a1)
+        .set    reorder
+
+.L_bn_sqr_words_return:
+        move    v0,zero
+        jr      ra
+
+.L_bn_sqr_words_tail:
+        dmultu  t0,t0
+        subu    a2,1
+        mflo    t1
+        mfhi    t0
+        sd      t1,0(a0)
+        sd      t0,8(a0)
+        beqz    a2,.L_bn_sqr_words_return
+
+        ld      t0,8(a1)
+        dmultu  t0,t0
+        subu    a2,1
+        mflo    t1
+        mfhi    t0
+        sd      t1,16(a0)
+        sd      t0,24(a0)
+        beqz    a2,.L_bn_sqr_words_return
+
+        ld      t0,16(a1)
+        dmultu  t0,t0
+        mflo    t1
+        mfhi    t0
+        sd      t1,32(a0)
+        sd      t0,40(a0)
+        jr      ra
+END(bn_sqr_words)
+
+.align  5
+LEAF(bn_add_words)
+        .set    noreorder
+        bgtzl   a3,.L_bn_add_words_proceed
+        ld      t0,0(a1)
+        jr      ra
+        move    v0,zero
+        .set    reorder
+
+.L_bn_add_words_proceed:
+        li      MINUS4,-4
+        and     AT,a3,MINUS4
+        move    v0,zero
+        beqz    AT,.L_bn_add_words_tail
+
+.L_bn_add_words_loop:
+        ld      ta0,0(a2)
+        ld      t1,8(a1)
+        ld      ta1,8(a2)
+        ld      t2,16(a1)
+        ld      ta2,16(a2)
+        ld      t3,24(a1)
+        ld      ta3,24(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
+        sd      t2,-16(a0)
+        daddu   v0,t8
+        
+        daddu   ta3,t3
+        sltu    t9,ta3,t3
+        daddu   t3,ta3,v0
+        sltu    v0,t3,ta3
+        sd      t3,-8(a0)
+        daddu   v0,t9
+        
+        .set    noreorder
+        bgtzl   AT,.L_bn_add_words_loop
+        ld      t0,0(a1)
+
+        bnezl   a3,.L_bn_add_words_tail
+        ld      t0,0(a1)
+        .set    reorder
+
+.L_bn_add_words_return:
+        jr      ra
+
+.L_bn_add_words_tail:
+        ld      ta0,0(a2)
+        daddu   ta0,t0
+        subu    a3,1
+        sltu    t8,ta0,t0
+        daddu   t0,ta0,v0
+        sltu    v0,t0,ta0
+        sd      t0,0(a0)
+        daddu   v0,t8
+        beqz    a3,.L_bn_add_words_return
+
+        ld      t1,8(a1)
+        ld      ta1,8(a2)
+        daddu   ta1,t1
+        subu    a3,1
+        sltu    t9,ta1,t1
+        daddu   t1,ta1,v0
+        sltu    v0,t1,ta1
+        sd      t1,8(a0)
+        daddu   v0,t9
+        beqz    a3,.L_bn_add_words_return
+
+        ld      t2,16(a1)
+        ld      ta2,16(a2)
+        daddu   ta2,t2
+        sltu    t8,ta2,t2
+        daddu   t2,ta2,v0
+        sltu    v0,t2,ta2
+        sd      t2,16(a0)
+        daddu   v0,t8
+        jr      ra
+END(bn_add_words)
+
+.align  5
+LEAF(bn_sub_words)
+        .set    noreorder
+        bgtzl   a3,.L_bn_sub_words_proceed
+        ld      t0,0(a1)
+        jr      ra
+        move    v0,zero
+        .set    reorder
+
+.L_bn_sub_words_proceed:
+        li      MINUS4,-4
+        and     AT,a3,MINUS4
+        move    v0,zero
+        beqz    AT,.L_bn_sub_words_tail
+
+.L_bn_sub_words_loop:
+        ld      ta0,0(a2)
+        ld      t1,8(a1)
+        ld      ta1,8(a2)
+        ld      t2,16(a1)
+        ld      ta2,16(a2)
+        ld      t3,24(a1)
+        ld      ta3,24(a2)
+        sltu    t8,t0,ta0
+        dsubu   t0,ta0
+        subu    a3,4
+        dsubu   ta0,t0,v0
+        and     AT,a3,MINUS4
+        sd      ta0,0(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)
+
+
+        sltu    t8,t2,ta2
+        dsubu   t2,ta2
+        dsubu   ta2,t2,v0
+        PTR_ADD a2,32
+        sd      ta2,-16(a0)
+        MOVNZ   (t2,v0,t8)
+
+        sltu    t9,t3,ta3
+        dsubu   t3,ta3
+        dsubu   ta3,t3,v0
+        sd      ta3,-8(a0)
+        MOVNZ   (t3,v0,t9)
+
+        .set    noreorder
+        bgtzl   AT,.L_bn_sub_words_loop
+        ld      t0,0(a1)
+
+        bnezl   a3,.L_bn_sub_words_tail
+        ld      t0,0(a1)
+        .set    reorder
+
+.L_bn_sub_words_return:
+        jr      ra
+
+.L_bn_sub_words_tail:
+        ld      ta0,0(a2)
+        subu    a3,1
+        sltu    t8,t0,ta0
+        dsubu   t0,ta0
+        dsubu   ta0,t0,v0
+        MOVNZ   (t0,v0,t8)
+        sd      ta0,0(a0)
+        beqz    a3,.L_bn_sub_words_return
+
+        ld      t1,8(a1)
+        subu    a3,1
+        ld      ta1,8(a2)
+        sltu    t9,t1,ta1
+        dsubu   t1,ta1
+        dsubu   ta1,t1,v0
+        MOVNZ   (t1,v0,t9)
+        sd      ta1,8(a0)
+        beqz    a3,.L_bn_sub_words_return
+
+        ld      t2,16(a1)
+        ld      ta2,16(a2)
+        sltu    t8,t2,ta2
+        dsubu   t2,ta2
+        dsubu   ta2,t2,v0
+        MOVNZ   (t2,v0,t8)
+        sd      ta2,16(a0)
+        jr      ra
+END(bn_sub_words)
+
+#undef  MINUS4
+
+.align  5
+LEAF(bn_div_words)
+        .set    noreorder
+        bnezl   a2,.L_bn_div_words_proceed
+        move    v1,zero
+        jr      ra
+        li      v0,-1           /* I'd rather signal div-by-zero
+                                 * which can be done with 'break 7' */
+
+.L_bn_div_words_proceed:
+        bltz    a2,.L_bn_div_words_body
+        move    t9,v1
+        dsll    a2,1
+        bgtz    a2,.-4
+        addu    t9,1
+
+        .set    reorder
+        negu    t1,t9
+        li      t2,-1
+        dsll    t2,t1
+        and     t2,a0
+        dsrl    AT,a1,t1
+        .set    noreorder
+        bnezl   t2,.+8
+        break   6               /* signal overflow */
+        .set    reorder
+        dsll    a0,t9
+        dsll    a1,t9
+        or      a0,AT
+
+#define QT      ta0
+#define HH      ta1
+#define DH      v1
+.L_bn_div_words_body:
+        dsrl    DH,a2,32
+        sgeu    AT,a0,a2
+        .set    noreorder
+        bnezl   AT,.+8
+        dsubu   a0,a2
+        .set    reorder
+
+        li      QT,-1
+        dsrl    HH,a0,32
+        dsrl    QT,32   /* q=0xffffffff */
+        beq     DH,HH,.L_bn_div_words_skip_div1
+        ddivu   zero,a0,DH
+        mflo    QT
+.L_bn_div_words_skip_div1:
+        dmultu  a2,QT
+        dsll    t3,a0,32
+        dsrl    AT,a1,32
+        or      t3,AT
+        mflo    t0
+        mfhi    t1
+.L_bn_div_words_inner_loop1:
+        sltu    t2,t3,t0
+        seq     t8,HH,t1
+        sltu    AT,HH,t1
+        and     t2,t8
+        or      AT,t2
+        .set    noreorder
+        beqz    AT,.L_bn_div_words_inner_loop1_done
+        sltu    t2,t0,a2
+        .set    reorder
+        dsubu   QT,1
+        dsubu   t0,a2
+        dsubu   t1,t2
+        b       .L_bn_div_words_inner_loop1
+.L_bn_div_words_inner_loop1_done:       
+
+        dsll    a1,32
+        dsubu   a0,t3,t0
+        dsll    v0,QT,32
+
+        li      QT,-1
+        dsrl    HH,a0,32
+        dsrl    QT,32   /* q=0xffffffff */
+        beq     DH,HH,.L_bn_div_words_skip_div2
+        ddivu   zero,a0,DH
+        mflo    QT
+.L_bn_div_words_skip_div2:
+        dmultu  a2,QT
+        dsll    t3,a0,32
+        dsrl    AT,a1,32
+        or      t3,AT
+        mflo    t0
+        mfhi    t1
+.L_bn_div_words_inner_loop2:
+        sltu    t2,t3,t0
+        seq     t8,HH,t1
+        sltu    AT,HH,t1
+        and     t2,t8
+        or      AT,t2
+        .set    noreorder
+        beqz    AT,.L_bn_div_words_inner_loop2_done
+        sltu    t2,t0,a2
+        .set    reorder
+        dsubu   QT,1
+        dsubu   t0,a2
+        dsubu   t1,t2
+        b       .L_bn_div_words_inner_loop2
+.L_bn_div_words_inner_loop2_done:       
+
+        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
+#define a_3     t3
+#define b_0     ta0
+#define b_1     ta1
+#define b_2     ta2
+#define b_3     ta3
+
+#define a_4     s0
+#define a_5     s2
+#define a_6     s4
+#define a_7     a1      /* once we load a[7] we don't need a anymore */
+#define b_4     s1
+#define b_5     s3
+#define b_6     s5
+#define b_7     a2      /* once we load b[7] we don't need b anymore */
+
+#define t_1     t8
+#define t_2     t9
+
+#define c_1     v0
+#define c_2     v1
+#define c_3     a3
+
+#define FRAME_SIZE      48
+
+.align  5
+LEAF(bn_mul_comba8)
+        .set    noreorder
+        PTR_SUB sp,FRAME_SIZE
+        .frame  sp,64,ra
+        .set    reorder
+        ld      a_0,0(a1)       /* If compiled with -mips3 option on
+                                 * R5000 box assembler barks on this
+                                 * line with "shouldn't have mult/div
+                                 * as last instruction in bb (R10K
+                                 * bug)" warning. If anybody out there
+                                 * has a clue about how to circumvent
+                                 * this do send me a note.
+                                 *              <appro@fy.chalmers.se>
+                                 */
+        ld      b_0,0(a2)
+        ld      a_1,8(a1)
+        ld      a_2,16(a1)
+        ld      a_3,24(a1)
+        ld      b_1,8(a2)
+        ld      b_2,16(a2)
+        ld      b_3,24(a2)
+        dmultu  a_0,b_0         /* mul_add_c(a[0],b[0],c1,c2,c3); */
+        sd      s0,0(sp)
+        sd      s1,8(sp)
+        sd      s2,16(sp)
+        sd      s3,24(sp)
+        sd      s4,32(sp)
+        sd      s5,40(sp)
+        mflo    c_1
+        mfhi    c_2
+
+        dmultu  a_0,b_1         /* mul_add_c(a[0],b[1],c2,c3,c1); */
+        ld      a_4,32(a1)
+        ld      a_5,40(a1)
+        ld      a_6,48(a1)
+        ld      a_7,56(a1)
+        ld      b_4,32(a2)
+        ld      b_5,40(a2)
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   c_3,t_2,AT
+        dmultu  a_1,b_0         /* mul_add_c(a[1],b[0],c2,c3,c1); */
+        ld      b_6,48(a2)
+        ld      b_7,56(a2)
+        sd      c_1,0(a0)       /* r[0]=c1; */
+        mflo    t_1
+        mfhi    t_2
+        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)       /* r[1]=c2; */
+
+        dmultu  a_2,b_0         /* mul_add_c(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   t_2,AT
+        daddu   c_1,t_2
+        dmultu  a_1,b_1         /* mul_add_c(a[1],b[1],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        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
+        dmultu  a_0,b_2         /* mul_add_c(a[0],b[2],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        sd      c_3,16(a0)      /* r[2]=c3; */
+
+        dmultu  a_0,b_3         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_2,b_1         /* mul_add_c(a[2],b[1],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        dmultu  a_3,b_0         /* mul_add_c(a[3],b[0],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        sd      c_1,24(a0)      /* r[3]=c1; */
+
+        dmultu  a_4,b_0         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_2,b_2         /* mul_add_c(a[2],b[2],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   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
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        dmultu  a_0,b_4         /* mul_add_c(a[0],b[4],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        sd      c_2,32(a0)      /* r[4]=c2; */
+
+        dmultu  a_0,b_5         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_2,b_3         /* mul_add_c(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   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        dmultu  a_3,b_2         /* mul_add_c(a[3],b[2],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        dmultu  a_4,b_1         /* mul_add_c(a[4],b[1],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        dmultu  a_5,b_0         /* mul_add_c(a[5],b[0],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        sd      c_3,40(a0)      /* r[5]=c3; */
+
+        dmultu  a_6,b_0         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_4,b_2         /* mul_add_c(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   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        dmultu  a_3,b_3         /* mul_add_c(a[3],b[3],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        dmultu  a_2,b_4         /* mul_add_c(a[2],b[4],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        dmultu  a_1,b_5         /* mul_add_c(a[1],b[5],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        dmultu  a_0,b_6         /* mul_add_c(a[0],b[6],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        sd      c_1,48(a0)      /* r[6]=c1; */
+
+        dmultu  a_0,b_7         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_2,b_5         /* mul_add_c(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   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        dmultu  a_3,b_4         /* mul_add_c(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   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        dmultu  a_4,b_3         /* mul_add_c(a[4],b[3],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        dmultu  a_5,b_2         /* mul_add_c(a[5],b[2],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        dmultu  a_6,b_1         /* mul_add_c(a[6],b[1],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        dmultu  a_7,b_0         /* mul_add_c(a[7],b[0],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        sd      c_2,56(a0)      /* r[7]=c2; */
+
+        dmultu  a_7,b_1         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_5,b_3         /* mul_add_c(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   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        dmultu  a_4,b_4         /* mul_add_c(a[4],b[4],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        dmultu  a_3,b_5         /* mul_add_c(a[3],b[5],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        dmultu  a_2,b_6         /* mul_add_c(a[2],b[6],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        dmultu  a_1,b_7         /* mul_add_c(a[1],b[7],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        sd      c_3,64(a0)      /* r[8]=c3; */
+
+        dmultu  a_2,b_7         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_4,b_5         /* mul_add_c(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   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        dmultu  a_5,b_4         /* mul_add_c(a[5],b[4],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        dmultu  a_6,b_3         /* mul_add_c(a[6],b[3],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        dmultu  a_7,b_2         /* mul_add_c(a[7],b[2],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        sd      c_1,72(a0)      /* r[9]=c1; */
+
+        dmultu  a_7,b_3         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_5,b_5         /* mul_add_c(a[5],b[5],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        dmultu  a_4,b_6         /* mul_add_c(a[4],b[6],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        dmultu  a_3,b_7         /* mul_add_c(a[3],b[7],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        sd      c_2,80(a0)      /* r[10]=c2; */
+
+        dmultu  a_4,b_7         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_6,b_5         /* mul_add_c(a[6],b[5],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        dmultu  a_7,b_4         /* mul_add_c(a[7],b[4],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        sd      c_3,88(a0)      /* r[11]=c3; */
+
+        dmultu  a_7,b_5         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_5,b_7         /* mul_add_c(a[5],b[7],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        sd      c_1,96(a0)      /* r[12]=c1; */
+
+        dmultu  a_6,b_7         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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_2,104(a0)     /* r[13]=c2; */
+
+        dmultu  a_7,b_7         /* mul_add_c(a[7],b[7],c3,c1,c2); */
+        ld      s0,0(sp)
+        ld      s1,8(sp)
+        ld      s2,16(sp)
+        ld      s3,24(sp)
+        ld      s4,32(sp)
+        ld      s5,40(sp)
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sd      c_3,112(a0)     /* r[14]=c3; */
+        sd      c_1,120(a0)     /* r[15]=c1; */
+
+        PTR_ADD sp,FRAME_SIZE
+
+        jr      ra
+END(bn_mul_comba8)
+
+.align  5
+LEAF(bn_mul_comba4)
+        .set    reorder
+        ld      a_0,0(a1)
+        ld      b_0,0(a2)
+        ld      a_1,8(a1)
+        ld      a_2,16(a1)
+        dmultu  a_0,b_0         /* mul_add_c(a[0],b[0],c1,c2,c3); */
+        ld      a_3,24(a1)
+        ld      b_1,8(a2)
+        ld      b_2,16(a2)
+        ld      b_3,24(a2)
+        mflo    c_1
+        mfhi    c_2
+        sd      c_1,0(a0)
+
+        dmultu  a_0,b_1         /* mul_add_c(a[0],b[1],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   c_3,t_2,AT
+        dmultu  a_1,b_0         /* mul_add_c(a[1],b[0],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        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,b_0         /* mul_add_c(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   t_2,AT
+        daddu   c_1,t_2
+        dmultu  a_1,b_1         /* mul_add_c(a[1],b[1],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        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
+        dmultu  a_0,b_2         /* mul_add_c(a[0],b[2],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        sd      c_3,16(a0)
+
+        dmultu  a_0,b_3         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_2,b_1         /* mul_add_c(a[2],b[1],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        dmultu  a_3,b_0         /* mul_add_c(a[3],b[0],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        sd      c_1,24(a0)
+
+        dmultu  a_3,b_1         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        dmultu  a_1,b_3         /* mul_add_c(a[1],b[3],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        sd      c_2,32(a0)
+
+        dmultu  a_2,b_3         /* mul_add_c(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   t_2,AT
+        daddu   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
+        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
+        sd      c_3,40(a0)
+
+        dmultu  a_3,b_3         /* mul_add_c(a[3],b[3],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sd      c_1,48(a0)
+        sd      c_2,56(a0)
+
+        jr      ra
+END(bn_mul_comba4)
+
+#undef  a_4
+#undef  a_5
+#undef  a_6
+#undef  a_7
+#define a_4     b_0
+#define a_5     b_1
+#define a_6     b_2
+#define a_7     b_3
+
+.align  5
+LEAF(bn_sqr_comba8)
+        .set    reorder
+        ld      a_0,0(a1)
+        ld      a_1,8(a1)
+        ld      a_2,16(a1)
+        ld      a_3,24(a1)
+
+        dmultu  a_0,a_0         /* mul_add_c(a[0],b[0],c1,c2,c3); */
+        ld      a_4,32(a1)
+        ld      a_5,40(a1)
+        ld      a_6,48(a1)
+        ld      a_7,56(a1)
+        mflo    c_1
+        mfhi    c_2
+        sd      c_1,0(a0)
+
+        dmultu  a_0,a_1         /* mul_add_c2(a[0],b[1],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        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
+        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
+        dmultu  a_1,a_1         /* mul_add_c(a[1],b[1],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        sd      c_3,16(a0)
+
+        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
+        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
+        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
+        daddu   c_3,AT
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        sd      c_1,24(a0)
+
+        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
+        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
+        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
+        daddu   c_1,AT
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   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
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        sd      c_2,32(a0)
+
+        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
+        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
+        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
+        daddu   c_2,AT
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        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
+        daddu   c_2,AT
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        sd      c_3,40(a0)
+
+        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
+        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
+        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
+        daddu   c_3,AT
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        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
+        daddu   c_3,AT
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        dmultu  a_3,a_3         /* mul_add_c(a[3],b[3],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        sd      c_1,48(a0)
+
+        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
+        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
+        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
+        daddu   c_1,AT
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        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
+        daddu   c_1,AT
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        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
+        daddu   c_1,AT
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        sd      c_2,56(a0)
+
+        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
+        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
+        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
+        daddu   c_2,AT
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        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
+        daddu   c_2,AT
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        dmultu  a_4,a_4         /* mul_add_c(a[4],b[4],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        sd      c_3,64(a0)
+
+        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
+        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
+        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
+        daddu   c_3,AT
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        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
+        daddu   c_3,AT
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        sd      c_1,72(a0)
+
+        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
+        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
+        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
+        daddu   c_1,AT
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        dmultu  a_5,a_5         /* mul_add_c(a[5],b[5],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        sd      c_2,80(a0)
+
+        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
+        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
+        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
+        daddu   c_2,AT
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        sd      c_3,88(a0)
+
+        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
+        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
+        dmultu  a_6,a_6         /* mul_add_c(a[6],b[6],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        sd      c_1,96(a0)
+
+        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
+        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_2,104(a0)
+
+        dmultu  a_7,a_7         /* mul_add_c(a[7],b[7],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sd      c_3,112(a0)
+        sd      c_1,120(a0)
+
+        jr      ra
+END(bn_sqr_comba8)
+
+.align  5
+LEAF(bn_sqr_comba4)
+        .set    reorder
+        ld      a_0,0(a1)
+        ld      a_1,8(a1)
+        ld      a_2,16(a1)
+        ld      a_3,24(a1)
+        dmultu  a_0,a_0         /* mul_add_c(a[0],b[0],c1,c2,c3); */
+        mflo    c_1
+        mfhi    c_2
+        sd      c_1,0(a0)
+
+        dmultu  a_0,a_1         /* mul_add_c2(a[0],b[1],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        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
+        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
+        dmultu  a_1,a_1         /* mul_add_c(a[1],b[1],c3,c1,c2); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_3,t_1
+        sltu    AT,c_3,t_1
+        daddu   t_2,AT
+        daddu   c_1,t_2
+        sltu    AT,c_1,t_2
+        daddu   c_2,AT
+        sd      c_3,16(a0)
+
+        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
+        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
+        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
+        daddu   c_3,AT
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sltu    AT,c_2,t_2
+        daddu   c_3,AT
+        sd      c_1,24(a0)
+
+        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
+        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
+        dmultu  a_2,a_2         /* mul_add_c(a[2],b[2],c2,c3,c1); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_2,t_1
+        sltu    AT,c_2,t_1
+        daddu   t_2,AT
+        daddu   c_3,t_2
+        sltu    AT,c_3,t_2
+        daddu   c_1,AT
+        sd      c_2,32(a0)
+
+        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
+        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
+        sd      c_3,40(a0)
+
+        dmultu  a_3,a_3         /* mul_add_c(a[3],b[3],c1,c2,c3); */
+        mflo    t_1
+        mfhi    t_2
+        daddu   c_1,t_1
+        sltu    AT,c_1,t_1
+        daddu   t_2,AT
+        daddu   c_2,t_2
+        sd      c_1,48(a0)
+        sd      c_2,56(a0)
+
+        jr      ra
+END(bn_sqr_comba4)
diff --git a/src/lib/libcrypto/bn/asm/sparc.s b/src/lib/libcrypto/bn/asm/sparc.s
deleted file mode 100644
index f9e533caa8..0000000000
--- a/src/lib/libcrypto/bn/asm/sparc.s
+++ /dev/null
@@ -1,462 +0,0 @@
-        .file   "bn_mulw.c"
-gcc2_compiled.:
-.section        ".text"
-        .align 4
-        .global bn_mul_add_words
-        .type    bn_mul_add_words,#function
-        .proc   016
-bn_mul_add_words:
-        !#PROLOGUE# 0
-        save %sp,-112,%sp
-        !#PROLOGUE# 1
-        mov %i0,%o0
-        mov %i1,%o2
-        mov %i2,%g1
-        mov %i3,%o1
-        mov 0,%i4
-        add %o0,12,%g4
-        add %o2,12,%o7
-.LL2:
-        mov %i4,%i3
-        mov 0,%i2
-        ld [%o0],%g2
-        mov %g2,%i1
-        ld [%o2],%g2
-        mov 0,%i0
-        umul %o1,%g2,%g3
-        rd %y,%g2
-        addcc %g3,%i1,%g3
-        addx %g2,%i0,%g2
-        addcc %g3,%i3,%g3
-        addx %g2,%i2,%g2
-        st %g3,[%o0]
-        mov %g2,%i5
-        mov 0,%i4
-        addcc %g1,-1,%g1
-        be .LL3
-        mov %i5,%i4
-        mov %i4,%i3
-        mov 0,%i2
-        ld [%g4-8],%g2
-        mov %g2,%i1
-        ld [%o7-8],%g2
-        mov 0,%i0
-        umul %o1,%g2,%g3
-        rd %y,%g2
-        addcc %g3,%i1,%g3
-        addx %g2,%i0,%g2
-        addcc %g3,%i3,%g3
-        addx %g2,%i2,%g2
-        st %g3,[%g4-8]
-        mov %g2,%i5
-        mov 0,%i4
-        addcc %g1,-1,%g1
-        be .LL3
-        mov %i5,%i4
-        mov %i4,%i3
-        mov 0,%i2
-        ld [%g4-4],%g2
-        mov %g2,%i1
-        ld [%o7-4],%g2
-        mov 0,%i0
-        umul %o1,%g2,%g3
-        rd %y,%g2
-        addcc %g3,%i1,%g3
-        addx %g2,%i0,%g2
-        addcc %g3,%i3,%g3
-        addx %g2,%i2,%g2
-        st %g3,[%g4-4]
-        mov %g2,%i5
-        mov 0,%i4
-        addcc %g1,-1,%g1
-        be .LL3
-        mov %i5,%i4
-        mov %i4,%i3
-        mov 0,%i2
-        ld [%g4],%g2
-        mov %g2,%i1
-        ld [%o7],%g2
-        mov 0,%i0
-        umul %o1,%g2,%g3
-        rd %y,%g2
-        addcc %g3,%i1,%g3
-        addx %g2,%i0,%g2
-        addcc %g3,%i3,%g3
-        addx %g2,%i2,%g2
-        st %g3,[%g4]
-        mov %g2,%i5
-        mov 0,%i4
-        addcc %g1,-1,%g1
-        be .LL3
-        mov %i5,%i4
-        add %o7,16,%o7
-        add %o2,16,%o2
-        add %g4,16,%g4
-        b .LL2
-        add %o0,16,%o0
-.LL3:
-        ret
-        restore %g0,%i4,%o0
-.LLfe1:
-        .size    bn_mul_add_words,.LLfe1-bn_mul_add_words
-        .align 4
-        .global bn_mul_words
-        .type    bn_mul_words,#function
-        .proc   016
-bn_mul_words:
-        !#PROLOGUE# 0
-        save %sp,-112,%sp
-        !#PROLOGUE# 1
-        mov %i0,%o7
-        mov %i1,%o0
-        mov %i2,%i4
-        mov %i3,%g4
-        mov 0,%i0
-        add %o7,12,%g1
-        add %o0,12,%i5
-.LL18:
-        mov %i0,%g3
-        mov 0,%g2
-        ld [%o0],%i2
-        umul %g4,%i2,%i3
-        rd %y,%i2
-        addcc %i3,%g3,%i3
-        addx %i2,%g2,%i2
-        st %i3,[%o7]
-        mov %i2,%i1
-        mov 0,%i0
-        addcc %i4,-1,%i4
-        be .LL19
-        mov %i1,%i0
-        mov %i0,%g3
-        mov 0,%g2
-        ld [%i5-8],%i2
-        umul %g4,%i2,%i3
-        rd %y,%i2
-        addcc %i3,%g3,%i3
-        addx %i2,%g2,%i2
-        st %i3,[%g1-8]
-        mov %i2,%i1
-        mov 0,%i0
-        addcc %i4,-1,%i4
-        be .LL19
-        mov %i1,%i0
-        mov %i0,%g3
-        mov 0,%g2
-        ld [%i5-4],%i2
-        umul %g4,%i2,%i3
-        rd %y,%i2
-        addcc %i3,%g3,%i3
-        addx %i2,%g2,%i2
-        st %i3,[%g1-4]
-        mov %i2,%i1
-        mov 0,%i0
-        addcc %i4,-1,%i4
-        be .LL19
-        mov %i1,%i0
-        mov %i0,%g3
-        mov 0,%g2
-        ld [%i5],%i2
-        umul %g4,%i2,%i3
-        rd %y,%i2
-        addcc %i3,%g3,%i3
-        addx %i2,%g2,%i2
-        st %i3,[%g1]
-        mov %i2,%i1
-        mov 0,%i0
-        addcc %i4,-1,%i4
-        be .LL19
-        mov %i1,%i0
-        add %i5,16,%i5
-        add %o0,16,%o0
-        add %g1,16,%g1
-        b .LL18
-        add %o7,16,%o7
-.LL19:
-        ret
-        restore
-.LLfe2:
-        .size    bn_mul_words,.LLfe2-bn_mul_words
-        .align 4
-        .global bn_sqr_words
-        .type    bn_sqr_words,#function
-        .proc   020
-bn_sqr_words:
-        !#PROLOGUE# 0
-        !#PROLOGUE# 1
-        mov %o0,%g4
-        add %g4,28,%o3
-        add %o1,12,%g1
-.LL34:
-        ld [%o1],%o0
-        addcc %o2,-1,%o2
-        umul %o0,%o0,%o5
-        rd %y,%o4
-        st %o5,[%g4]
-        mov %o4,%g3
-        mov 0,%g2
-        be .LL35
-        st %g3,[%o3-24]
-        ld [%g1-8],%o0
-        addcc %o2,-1,%o2
-        umul %o0,%o0,%o5
-        rd %y,%o4
-        st %o5,[%o3-20]
-        mov %o4,%g3
-        mov 0,%g2
-        be .LL35
-        st %g3,[%o3-16]
-        ld [%g1-4],%o0
-        addcc %o2,-1,%o2
-        umul %o0,%o0,%o5
-        rd %y,%o4
-        st %o5,[%o3-12]
-        mov %o4,%g3
-        mov 0,%g2
-        be .LL35
-        st %g3,[%o3-8]
-        ld [%g1],%o0
-        addcc %o2,-1,%o2
-        umul %o0,%o0,%o5
-        rd %y,%o4
-        st %o5,[%o3-4]
-        mov %o4,%g3
-        mov 0,%g2
-        be .LL35
-        st %g3,[%o3]
-        add %g1,16,%g1
-        add %o1,16,%o1
-        add %o3,32,%o3
-        b .LL34
-        add %g4,32,%g4
-.LL35:
-        retl
-        nop
-.LLfe3:
-        .size    bn_sqr_words,.LLfe3-bn_sqr_words
-        .align 4
-        .global bn_add_words
-        .type    bn_add_words,#function
-        .proc   016
-bn_add_words:
-        !#PROLOGUE# 0
-        save %sp,-112,%sp
-        !#PROLOGUE# 1
-        mov %i0,%o2
-        mov %i1,%o3
-        mov %i2,%o4
-        mov %i3,%i5
-        mov 0,%o0
-        mov 0,%o1
-        add %o2,12,%o7
-        add %o4,12,%g4
-        b .LL42
-        add %o3,12,%g1
-.LL45:
-        add %i5,-1,%i5
-        mov %i4,%g3
-        ld [%g4-8],%i4
-        mov 0,%g2
-        mov %i4,%i1
-        mov 0,%i0
-        addcc %g3,%i1,%g3
-        addx %g2,%i0,%g2
-        addcc %o1,%g3,%o1
-        addx %o0,%g2,%o0
-        st %o1,[%o7-8]
-        mov %o0,%i3
-        mov 0,%i2
-        mov %i2,%o0
-        mov %i3,%o1
-        cmp %i5,0
-        ble .LL43
-        add %i5,-1,%i5
-        ld [%g1-4],%i4
-        mov %i4,%g3
-        ld [%g4-4],%i4
-        mov 0,%g2
-        mov %i4,%i1
-        mov 0,%i0
-        addcc %g3,%i1,%g3
-        addx %g2,%i0,%g2
-        addcc %o1,%g3,%o1
-        addx %o0,%g2,%o0
-        st %o1,[%o7-4]
-        mov %o0,%i3
-        mov 0,%i2
-        mov %i2,%o0
-        mov %i3,%o1
-        cmp %i5,0
-        ble .LL43
-        add %i5,-1,%i5
-        ld [%g1],%i4
-        mov %i4,%g3
-        ld [%g4],%i4
-        mov 0,%g2
-        mov %i4,%i1
-        mov 0,%i0
-        addcc %g3,%i1,%g3
-        addx %g2,%i0,%g2
-        addcc %o1,%g3,%o1
-        addx %o0,%g2,%o0
-        st %o1,[%o7]
-        mov %o0,%i3
-        mov 0,%i2
-        mov %i2,%o0
-        mov %i3,%o1
-        cmp %i5,0
-        ble .LL43
-        add %g1,16,%g1
-        add %o3,16,%o3
-        add %g4,16,%g4
-        add %o4,16,%o4
-        add %o7,16,%o7
-        add %o2,16,%o2
-.LL42:
-        ld [%o3],%i4
-        add %i5,-1,%i5
-        mov %i4,%g3
-        ld [%o4],%i4
-        mov 0,%g2
-        mov %i4,%i1
-        mov 0,%i0
-        addcc %g3,%i1,%g3
-        addx %g2,%i0,%g2
-        addcc %o1,%g3,%o1
-        addx %o0,%g2,%o0
-        st %o1,[%o2]
-        mov %o0,%i3
-        mov 0,%i2
-        mov %i2,%o0
-        mov %i3,%o1
-        cmp %i5,0
-        bg,a .LL45
-        ld [%g1-8],%i4
-.LL43:
-        ret
-        restore %g0,%o1,%o0
-.LLfe4:
-        .size    bn_add_words,.LLfe4-bn_add_words
-.section        ".rodata"
-        .align 8
-.LLC0:
-        .asciz  "Division would overflow (%d)\n"
-.section        ".text"
-        .align 4
-        .global bn_div64
-        .type    bn_div64,#function
-        .proc   016
-bn_div64:
-        !#PROLOGUE# 0
-        save %sp,-112,%sp
-        !#PROLOGUE# 1
-        mov 0,%l1
-        cmp %i2,0
-        bne .LL51
-        mov 2,%l0
-        b .LL68
-        mov -1,%i0
-.LL51:
-        call BN_num_bits_word,0
-        mov %i2,%o0
-        mov %o0,%o2
-        cmp %o2,32
-        be .LL52
-        mov 1,%o0
-        sll %o0,%o2,%o0
-        cmp %i0,%o0
-        bleu .LL69
-        mov 32,%o0
-        sethi %hi(__iob+32),%o0
-        or %o0,%lo(__iob+32),%o0
-        sethi %hi(.LLC0),%o1
-        call fprintf,0
-        or %o1,%lo(.LLC0),%o1
-        call abort,0
-        nop
-.LL52:
-        mov 32,%o0
-.LL69:
-        cmp %i0,%i2
-        blu .LL53
-        sub %o0,%o2,%o2
-        sub %i0,%i2,%i0
-.LL53:
-        cmp %o2,0
-        be .LL54
-        sll %i0,%o2,%o1
-        sll %i2,%o2,%i2
-        sub %o0,%o2,%o0
-        srl %i1,%o0,%o0
-        or %o1,%o0,%i0
-        sll %i1,%o2,%i1
-.LL54:
-        srl %i2,16,%g2
-        sethi %hi(65535),%o0
-        or %o0,%lo(65535),%o1
-        and %i2,%o1,%g3
-        mov %o0,%g4
-        sethi %hi(-65536),%o7
-        mov %o1,%g1
-.LL55:
-        srl %i0,16,%o0
-        cmp %o0,%g2
-        be .LL59
-        or %g4,%lo(65535),%o3
-        wr %g0,%g0,%y
-        nop
-        nop
-        nop
-        udiv %i0,%g2,%o3
-.LL59:
-        and %i1,%o7,%o0
-        srl %o0,16,%o5
-        smul %o3,%g3,%o4
-        smul %o3,%g2,%o2
-.LL60:
-        sub %i0,%o2,%o1
-        andcc %o1,%o7,%g0
-        bne .LL61
-        sll %o1,16,%o0
-        add %o0,%o5,%o0
-        cmp %o4,%o0
-        bleu .LL61
-        sub %o4,%g3,%o4
-        sub %o2,%g2,%o2
-        b .LL60
-        add %o3,-1,%o3
-.LL61:
-        smul %o3,%g2,%o2
-        smul %o3,%g3,%o0
-        srl %o0,16,%o1
-        sll %o0,16,%o0
-        and %o0,%o7,%o0
-        cmp %i1,%o0
-        bgeu .LL65
-        add %o2,%o1,%o2
-        add %o2,1,%o2
-.LL65:
-        cmp %i0,%o2
-        bgeu .LL66
-        sub %i1,%o0,%i1
-        add %i0,%i2,%i0
-        add %o3,-1,%o3
-.LL66:
-        addcc %l0,-1,%l0
-        be .LL56
-        sub %i0,%o2,%i0
-        sll %o3,16,%l1
-        sll %i0,16,%o0
-        srl %i1,16,%o1
-        or %o0,%o1,%i0
-        and %i1,%g1,%o0
-        b .LL55
-        sll %o0,16,%i1
-.LL56:
-        or %l1,%o3,%i0
-.LL68:
-        ret
-        restore
-.LLfe5:
-        .size    bn_div64,.LLfe5-bn_div64
-        .ident  "GCC: (GNU) 2.7.2.3"
diff --git a/src/lib/libcrypto/bn/asm/sparcv8.S b/src/lib/libcrypto/bn/asm/sparcv8.S
new file mode 100644
index 0000000000..88c5dc480a
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/sparcv8.S
@@ -0,0 +1,1458 @@
+.ident  "sparcv8.s, Version 1.4"
+.ident  "SPARC v8 ISA artwork by Andy Polyakov <appro@fy.chalmers.se>"
+
+/*
+ * ====================================================================
+ * Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+ * project.
+ *
+ * Rights for redistribution and usage in source and binary forms are
+ * granted according to the OpenSSL license. Warranty of any kind is
+ * disclaimed.
+ * ====================================================================
+ */
+
+/*
+ * This is my modest contributon to OpenSSL project (see
+ * http://www.openssl.org/ for more information about it) and is
+ * a drop-in SuperSPARC ISA replacement for crypto/bn/bn_asm.c
+ * module. For updates see http://fy.chalmers.se/~appro/hpe/.
+ *
+ * See bn_asm.sparc.v8plus.S for more details.
+ */
+
+/*
+ * Revision history.
+ *
+ * 1.1  - new loop unrolling model(*);
+ * 1.2  - made gas friendly;
+ * 1.3  - fixed problem with /usr/ccs/lib/cpp;
+ * 1.4  - some retunes;
+ *
+ * (*)  see bn_asm.sparc.v8plus.S for details
+ */
+
+.section        ".text",#alloc,#execinstr
+.file           "bn_asm.sparc.v8.S"
+
+.align  32
+
+.global bn_mul_add_words
+/*
+ * BN_ULONG bn_mul_add_words(rp,ap,num,w)
+ * BN_ULONG *rp,*ap;
+ * int num;
+ * BN_ULONG w;
+ */
+bn_mul_add_words:
+        cmp     %o2,0
+        bg,a    .L_bn_mul_add_words_proceed
+        ld      [%o1],%g2
+        retl
+        clr     %o0
+
+.L_bn_mul_add_words_proceed:
+        andcc   %o2,-4,%g0
+        bz      .L_bn_mul_add_words_tail
+        clr     %o5
+
+.L_bn_mul_add_words_loop:
+        ld      [%o0],%o4
+        ld      [%o1+4],%g3
+        umul    %o3,%g2,%g2
+        rd      %y,%g1
+        addcc   %o4,%o5,%o4
+        addx    %g1,0,%g1
+        addcc   %o4,%g2,%o4
+        st      %o4,[%o0]
+        addx    %g1,0,%o5
+
+        ld      [%o0+4],%o4
+        ld      [%o1+8],%g2
+        umul    %o3,%g3,%g3
+        dec     4,%o2
+        rd      %y,%g1
+        addcc   %o4,%o5,%o4
+        addx    %g1,0,%g1
+        addcc   %o4,%g3,%o4
+        st      %o4,[%o0+4]
+        addx    %g1,0,%o5
+
+        ld      [%o0+8],%o4
+        ld      [%o1+12],%g3
+        umul    %o3,%g2,%g2
+        inc     16,%o1
+        rd      %y,%g1
+        addcc   %o4,%o5,%o4
+        addx    %g1,0,%g1
+        addcc   %o4,%g2,%o4
+        st      %o4,[%o0+8]
+        addx    %g1,0,%o5
+
+        ld      [%o0+12],%o4
+        umul    %o3,%g3,%g3
+        inc     16,%o0
+        rd      %y,%g1
+        addcc   %o4,%o5,%o4
+        addx    %g1,0,%g1
+        addcc   %o4,%g3,%o4
+        st      %o4,[%o0-4]
+        addx    %g1,0,%o5
+        andcc   %o2,-4,%g0
+        bnz,a   .L_bn_mul_add_words_loop
+        ld      [%o1],%g2
+
+        tst     %o2
+        bnz,a   .L_bn_mul_add_words_tail
+        ld      [%o1],%g2
+.L_bn_mul_add_words_return:
+        retl
+        mov     %o5,%o0
+        nop
+
+.L_bn_mul_add_words_tail:
+        ld      [%o0],%o4
+        umul    %o3,%g2,%g2
+        addcc   %o4,%o5,%o4
+        rd      %y,%g1
+        addx    %g1,0,%g1
+        addcc   %o4,%g2,%o4
+        addx    %g1,0,%o5
+        deccc   %o2
+        bz      .L_bn_mul_add_words_return
+        st      %o4,[%o0]
+
+        ld      [%o1+4],%g2
+        ld      [%o0+4],%o4
+        umul    %o3,%g2,%g2
+        rd      %y,%g1
+        addcc   %o4,%o5,%o4
+        addx    %g1,0,%g1
+        addcc   %o4,%g2,%o4
+        addx    %g1,0,%o5
+        deccc   %o2
+        bz      .L_bn_mul_add_words_return
+        st      %o4,[%o0+4]
+
+        ld      [%o1+8],%g2
+        ld      [%o0+8],%o4
+        umul    %o3,%g2,%g2
+        rd      %y,%g1
+        addcc   %o4,%o5,%o4
+        addx    %g1,0,%g1
+        addcc   %o4,%g2,%o4
+        st      %o4,[%o0+8]
+        retl
+        addx    %g1,0,%o0
+
+.type   bn_mul_add_words,#function
+.size   bn_mul_add_words,(.-bn_mul_add_words)
+
+.align  32
+
+.global bn_mul_words
+/*
+ * BN_ULONG bn_mul_words(rp,ap,num,w)
+ * BN_ULONG *rp,*ap;
+ * int num;
+ * BN_ULONG w;
+ */
+bn_mul_words:
+        cmp     %o2,0
+        bg,a    .L_bn_mul_words_proceeed
+        ld      [%o1],%g2
+        retl
+        clr     %o0
+
+.L_bn_mul_words_proceeed:
+        andcc   %o2,-4,%g0
+        bz      .L_bn_mul_words_tail
+        clr     %o5
+
+.L_bn_mul_words_loop:
+        ld      [%o1+4],%g3
+        umul    %o3,%g2,%g2
+        addcc   %g2,%o5,%g2
+        rd      %y,%g1
+        addx    %g1,0,%o5
+        st      %g2,[%o0]
+
+        ld      [%o1+8],%g2
+        umul    %o3,%g3,%g3
+        addcc   %g3,%o5,%g3
+        rd      %y,%g1
+        dec     4,%o2
+        addx    %g1,0,%o5
+        st      %g3,[%o0+4]
+
+        ld      [%o1+12],%g3
+        umul    %o3,%g2,%g2
+        addcc   %g2,%o5,%g2
+        rd      %y,%g1
+        inc     16,%o1
+        st      %g2,[%o0+8]
+        addx    %g1,0,%o5
+
+        umul    %o3,%g3,%g3
+        addcc   %g3,%o5,%g3
+        rd      %y,%g1
+        inc     16,%o0
+        addx    %g1,0,%o5
+        st      %g3,[%o0-4]
+        andcc   %o2,-4,%g0
+        nop
+        bnz,a   .L_bn_mul_words_loop
+        ld      [%o1],%g2
+
+        tst     %o2
+        bnz,a   .L_bn_mul_words_tail
+        ld      [%o1],%g2
+.L_bn_mul_words_return:
+        retl
+        mov     %o5,%o0
+        nop
+
+.L_bn_mul_words_tail:
+        umul    %o3,%g2,%g2
+        addcc   %g2,%o5,%g2
+        rd      %y,%g1
+        addx    %g1,0,%o5
+        deccc   %o2
+        bz      .L_bn_mul_words_return
+        st      %g2,[%o0]
+        nop
+
+        ld      [%o1+4],%g2
+        umul    %o3,%g2,%g2
+        addcc   %g2,%o5,%g2
+        rd      %y,%g1
+        addx    %g1,0,%o5
+        deccc   %o2
+        bz      .L_bn_mul_words_return
+        st      %g2,[%o0+4]
+
+        ld      [%o1+8],%g2
+        umul    %o3,%g2,%g2
+        addcc   %g2,%o5,%g2
+        rd      %y,%g1
+        st      %g2,[%o0+8]
+        retl
+        addx    %g1,0,%o0
+
+.type   bn_mul_words,#function
+.size   bn_mul_words,(.-bn_mul_words)
+
+.align  32
+.global bn_sqr_words
+/*
+ * void bn_sqr_words(r,a,n)
+ * BN_ULONG *r,*a;
+ * int n;
+ */
+bn_sqr_words:
+        cmp     %o2,0
+        bg,a    .L_bn_sqr_words_proceeed
+        ld      [%o1],%g2
+        retl
+        clr     %o0
+
+.L_bn_sqr_words_proceeed:
+        andcc   %o2,-4,%g0
+        bz      .L_bn_sqr_words_tail
+        clr     %o5
+
+.L_bn_sqr_words_loop:
+        ld      [%o1+4],%g3
+        umul    %g2,%g2,%o4
+        st      %o4,[%o0]
+        rd      %y,%o5
+        st      %o5,[%o0+4]
+
+        ld      [%o1+8],%g2
+        umul    %g3,%g3,%o4
+        dec     4,%o2
+        st      %o4,[%o0+8]
+        rd      %y,%o5
+        st      %o5,[%o0+12]
+        nop
+
+        ld      [%o1+12],%g3
+        umul    %g2,%g2,%o4
+        st      %o4,[%o0+16]
+        rd      %y,%o5
+        inc     16,%o1
+        st      %o5,[%o0+20]
+
+        umul    %g3,%g3,%o4
+        inc     32,%o0
+        st      %o4,[%o0-8]
+        rd      %y,%o5
+        st      %o5,[%o0-4]
+        andcc   %o2,-4,%g2
+        bnz,a   .L_bn_sqr_words_loop
+        ld      [%o1],%g2
+
+        tst     %o2
+        nop
+        bnz,a   .L_bn_sqr_words_tail
+        ld      [%o1],%g2
+.L_bn_sqr_words_return:
+        retl
+        clr     %o0
+
+.L_bn_sqr_words_tail:
+        umul    %g2,%g2,%o4
+        st      %o4,[%o0]
+        deccc   %o2
+        rd      %y,%o5
+        bz      .L_bn_sqr_words_return
+        st      %o5,[%o0+4]
+
+        ld      [%o1+4],%g2
+        umul    %g2,%g2,%o4
+        st      %o4,[%o0+8]
+        deccc   %o2
+        rd      %y,%o5
+        nop
+        bz      .L_bn_sqr_words_return
+        st      %o5,[%o0+12]
+
+        ld      [%o1+8],%g2
+        umul    %g2,%g2,%o4
+        st      %o4,[%o0+16]
+        rd      %y,%o5
+        st      %o5,[%o0+20]
+        retl
+        clr     %o0
+
+.type   bn_sqr_words,#function
+.size   bn_sqr_words,(.-bn_sqr_words)
+
+.align  32
+
+.global bn_div_words
+/*
+ * BN_ULONG bn_div_words(h,l,d)
+ * BN_ULONG h,l,d;
+ */
+bn_div_words:
+        wr      %o0,%y
+        udiv    %o1,%o2,%o0
+        retl
+        nop
+
+.type   bn_div_words,#function
+.size   bn_div_words,(.-bn_div_words)
+
+.align  32
+
+.global bn_add_words
+/*
+ * BN_ULONG bn_add_words(rp,ap,bp,n)
+ * BN_ULONG *rp,*ap,*bp;
+ * int n;
+ */
+bn_add_words:
+        cmp     %o3,0
+        bg,a    .L_bn_add_words_proceed
+        ld      [%o1],%o4
+        retl
+        clr     %o0
+
+.L_bn_add_words_proceed:
+        andcc   %o3,-4,%g0
+        bz      .L_bn_add_words_tail
+        clr     %g1
+        ba      .L_bn_add_words_warn_loop
+        addcc   %g0,0,%g0       ! clear carry flag
+
+.L_bn_add_words_loop:
+        ld      [%o1],%o4
+.L_bn_add_words_warn_loop:
+        ld      [%o2],%o5
+        ld      [%o1+4],%g3
+        ld      [%o2+4],%g4
+        dec     4,%o3
+        addxcc  %o5,%o4,%o5
+        st      %o5,[%o0]
+
+        ld      [%o1+8],%o4
+        ld      [%o2+8],%o5
+        inc     16,%o1
+        addxcc  %g3,%g4,%g3
+        st      %g3,[%o0+4]
+        
+        ld      [%o1-4],%g3
+        ld      [%o2+12],%g4
+        inc     16,%o2
+        addxcc  %o5,%o4,%o5
+        st      %o5,[%o0+8]
+
+        inc     16,%o0
+        addxcc  %g3,%g4,%g3
+        st      %g3,[%o0-4]
+        addx    %g0,0,%g1
+        andcc   %o3,-4,%g0
+        bnz,a   .L_bn_add_words_loop
+        addcc   %g1,-1,%g0
+
+        tst     %o3
+        bnz,a   .L_bn_add_words_tail
+        ld      [%o1],%o4
+.L_bn_add_words_return:
+        retl
+        mov     %g1,%o0
+
+.L_bn_add_words_tail:
+        addcc   %g1,-1,%g0
+        ld      [%o2],%o5
+        addxcc  %o5,%o4,%o5
+        addx    %g0,0,%g1
+        deccc   %o3
+        bz      .L_bn_add_words_return
+        st      %o5,[%o0]
+
+        ld      [%o1+4],%o4
+        addcc   %g1,-1,%g0
+        ld      [%o2+4],%o5
+        addxcc  %o5,%o4,%o5
+        addx    %g0,0,%g1
+        deccc   %o3
+        bz      .L_bn_add_words_return
+        st      %o5,[%o0+4]
+
+        ld      [%o1+8],%o4
+        addcc   %g1,-1,%g0
+        ld      [%o2+8],%o5
+        addxcc  %o5,%o4,%o5
+        st      %o5,[%o0+8]
+        retl
+        addx    %g0,0,%o0
+
+.type   bn_add_words,#function
+.size   bn_add_words,(.-bn_add_words)
+
+.align  32
+
+.global bn_sub_words
+/*
+ * BN_ULONG bn_sub_words(rp,ap,bp,n)
+ * BN_ULONG *rp,*ap,*bp;
+ * int n;
+ */
+bn_sub_words:
+        cmp     %o3,0
+        bg,a    .L_bn_sub_words_proceed
+        ld      [%o1],%o4
+        retl
+        clr     %o0
+
+.L_bn_sub_words_proceed:
+        andcc   %o3,-4,%g0
+        bz      .L_bn_sub_words_tail
+        clr     %g1
+        ba      .L_bn_sub_words_warm_loop
+        addcc   %g0,0,%g0       ! clear carry flag
+
+.L_bn_sub_words_loop:
+        ld      [%o1],%o4
+.L_bn_sub_words_warm_loop:
+        ld      [%o2],%o5
+        ld      [%o1+4],%g3
+        ld      [%o2+4],%g4
+        dec     4,%o3
+        subxcc  %o4,%o5,%o5
+        st      %o5,[%o0]
+
+        ld      [%o1+8],%o4
+        ld      [%o2+8],%o5
+        inc     16,%o1
+        subxcc  %g3,%g4,%g4
+        st      %g4,[%o0+4]
+        
+        ld      [%o1-4],%g3
+        ld      [%o2+12],%g4
+        inc     16,%o2
+        subxcc  %o4,%o5,%o5
+        st      %o5,[%o0+8]
+
+        inc     16,%o0
+        subxcc  %g3,%g4,%g4
+        st      %g4,[%o0-4]
+        addx    %g0,0,%g1
+        andcc   %o3,-4,%g0
+        bnz,a   .L_bn_sub_words_loop
+        addcc   %g1,-1,%g0
+
+        tst     %o3
+        nop
+        bnz,a   .L_bn_sub_words_tail
+        ld      [%o1],%o4
+.L_bn_sub_words_return:
+        retl
+        mov     %g1,%o0
+
+.L_bn_sub_words_tail:
+        addcc   %g1,-1,%g0
+        ld      [%o2],%o5
+        subxcc  %o4,%o5,%o5
+        addx    %g0,0,%g1
+        deccc   %o3
+        bz      .L_bn_sub_words_return
+        st      %o5,[%o0]
+        nop
+
+        ld      [%o1+4],%o4
+        addcc   %g1,-1,%g0
+        ld      [%o2+4],%o5
+        subxcc  %o4,%o5,%o5
+        addx    %g0,0,%g1
+        deccc   %o3
+        bz      .L_bn_sub_words_return
+        st      %o5,[%o0+4]
+
+        ld      [%o1+8],%o4
+        addcc   %g1,-1,%g0
+        ld      [%o2+8],%o5
+        subxcc  %o4,%o5,%o5
+        st      %o5,[%o0+8]
+        retl
+        addx    %g0,0,%o0
+
+.type   bn_sub_words,#function
+.size   bn_sub_words,(.-bn_sub_words)
+
+#define FRAME_SIZE      -96
+
+/*
+ * Here is register usage map for *all* routines below.
+ */
+#define t_1     %o0
+#define t_2     %o1
+#define c_1     %o2
+#define c_2     %o3
+#define c_3     %o4
+
+#define ap(I)   [%i1+4*I]
+#define bp(I)   [%i2+4*I]
+#define rp(I)   [%i0+4*I]
+
+#define a_0     %l0
+#define a_1     %l1
+#define a_2     %l2
+#define a_3     %l3
+#define a_4     %l4
+#define a_5     %l5
+#define a_6     %l6
+#define a_7     %l7
+
+#define b_0     %i3
+#define b_1     %i4
+#define b_2     %i5
+#define b_3     %o5
+#define b_4     %g1
+#define b_5     %g2
+#define b_6     %g3
+#define b_7     %g4
+
+.align  32
+.global bn_mul_comba8
+/*
+ * void bn_mul_comba8(r,a,b)
+ * BN_ULONG *r,*a,*b;
+ */
+bn_mul_comba8:
+        save    %sp,FRAME_SIZE,%sp
+        ld      ap(0),a_0
+        ld      bp(0),b_0
+        umul    a_0,b_0,c_1     !=!mul_add_c(a[0],b[0],c1,c2,c3);
+        ld      bp(1),b_1
+        rd      %y,c_2
+        st      c_1,rp(0)       !r[0]=c1;
+
+        umul    a_0,b_1,t_1     !=!mul_add_c(a[0],b[1],c2,c3,c1);
+        ld      ap(1),a_1
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  %g0,t_2,c_3     !=
+        addx    %g0,%g0,c_1
+        ld      ap(2),a_2
+        umul    a_1,b_0,t_1     !mul_add_c(a[1],b[0],c2,c3,c1);
+        addcc   c_2,t_1,c_2     !=
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        st      c_2,rp(1)       !r[1]=c2;
+        addx    c_1,%g0,c_1     !=
+
+        umul    a_2,b_0,t_1     !mul_add_c(a[2],b[0],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1     !=
+        addx    %g0,%g0,c_2
+        ld      bp(2),b_2
+        umul    a_1,b_1,t_1     !mul_add_c(a[1],b[1],c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        ld      bp(3),b_3
+        addx    c_2,%g0,c_2     !=
+        umul    a_0,b_2,t_1     !mul_add_c(a[0],b[2],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1     !=
+        addx    c_2,%g0,c_2
+        st      c_3,rp(2)       !r[2]=c3;
+
+        umul    a_0,b_3,t_1     !mul_add_c(a[0],b[3],c1,c2,c3);
+        addcc   c_1,t_1,c_1     !=
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    %g0,%g0,c_3
+        umul    a_1,b_2,t_1     !=!mul_add_c(a[1],b[2],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        ld      ap(3),a_3
+        umul    a_2,b_1,t_1     !mul_add_c(a[2],b[1],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2          !=
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3
+        ld      ap(4),a_4
+        umul    a_3,b_0,t_1     !mul_add_c(a[3],b[0],c1,c2,c3);!=
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        st      c_1,rp(3)       !r[3]=c1;
+
+        umul    a_4,b_0,t_1     !mul_add_c(a[4],b[0],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    %g0,%g0,c_1
+        umul    a_3,b_1,t_1     !mul_add_c(a[3],b[1],c2,c3,c1);
+        addcc   c_2,t_1,c_2     !=
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        umul    a_2,b_2,t_1     !=!mul_add_c(a[2],b[2],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1     !=
+        ld      bp(4),b_4
+        umul    a_1,b_3,t_1     !mul_add_c(a[1],b[3],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        ld      bp(5),b_5
+        umul    a_0,b_4,t_1     !=!mul_add_c(a[0],b[4],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1     !=
+        st      c_2,rp(4)       !r[4]=c2;
+
+        umul    a_0,b_5,t_1     !mul_add_c(a[0],b[5],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2          !=
+        addxcc  c_1,t_2,c_1
+        addx    %g0,%g0,c_2
+        umul    a_1,b_4,t_1     !mul_add_c(a[1],b[4],c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        umul    a_2,b_3,t_1     !=!mul_add_c(a[2],b[3],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2     !=
+        umul    a_3,b_2,t_1     !mul_add_c(a[3],b[2],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1     !=
+        addx    c_2,%g0,c_2
+        ld      ap(5),a_5
+        umul    a_4,b_1,t_1     !mul_add_c(a[4],b[1],c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        ld      ap(6),a_6
+        addx    c_2,%g0,c_2     !=
+        umul    a_5,b_0,t_1     !mul_add_c(a[5],b[0],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1     !=
+        addx    c_2,%g0,c_2
+        st      c_3,rp(5)       !r[5]=c3;
+
+        umul    a_6,b_0,t_1     !mul_add_c(a[6],b[0],c1,c2,c3);
+        addcc   c_1,t_1,c_1     !=
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    %g0,%g0,c_3
+        umul    a_5,b_1,t_1     !=!mul_add_c(a[5],b[1],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        umul    a_4,b_2,t_1     !mul_add_c(a[4],b[2],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    c_3,%g0,c_3
+        umul    a_3,b_3,t_1     !mul_add_c(a[3],b[3],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2          !=
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3
+        umul    a_2,b_4,t_1     !mul_add_c(a[2],b[4],c1,c2,c3);
+        addcc   c_1,t_1,c_1     !=
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        ld      bp(6),b_6
+        addx    c_3,%g0,c_3     !=
+        umul    a_1,b_5,t_1     !mul_add_c(a[1],b[5],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    c_3,%g0,c_3
+        ld      bp(7),b_7
+        umul    a_0,b_6,t_1     !mul_add_c(a[0],b[6],c1,c2,c3);
+        addcc   c_1,t_1,c_1     !=
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        st      c_1,rp(6)       !r[6]=c1;
+        addx    c_3,%g0,c_3     !=
+
+        umul    a_0,b_7,t_1     !mul_add_c(a[0],b[7],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    %g0,%g0,c_1
+        umul    a_1,b_6,t_1     !mul_add_c(a[1],b[6],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        umul    a_2,b_5,t_1     !mul_add_c(a[2],b[5],c2,c3,c1);
+        addcc   c_2,t_1,c_2     !=
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        umul    a_3,b_4,t_1     !=!mul_add_c(a[3],b[4],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1     !=
+        umul    a_4,b_3,t_1     !mul_add_c(a[4],b[3],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        umul    a_5,b_2,t_1     !mul_add_c(a[5],b[2],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        ld      ap(7),a_7
+        umul    a_6,b_1,t_1     !=!mul_add_c(a[6],b[1],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1     !=
+        umul    a_7,b_0,t_1     !mul_add_c(a[7],b[0],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        st      c_2,rp(7)       !r[7]=c2;
+
+        umul    a_7,b_1,t_1     !mul_add_c(a[7],b[1],c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    %g0,%g0,c_2
+        umul    a_6,b_2,t_1     !=!mul_add_c(a[6],b[2],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2     !=
+        umul    a_5,b_3,t_1     !mul_add_c(a[5],b[3],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1     !=
+        addx    c_2,%g0,c_2
+        umul    a_4,b_4,t_1     !mul_add_c(a[4],b[4],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2          !=
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        umul    a_3,b_5,t_1     !mul_add_c(a[3],b[5],c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        umul    a_2,b_6,t_1     !=!mul_add_c(a[2],b[6],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2     !=
+        umul    a_1,b_7,t_1     !mul_add_c(a[1],b[7],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1     !
+        addx    c_2,%g0,c_2
+        st      c_3,rp(8)       !r[8]=c3;
+
+        umul    a_2,b_7,t_1     !mul_add_c(a[2],b[7],c1,c2,c3);
+        addcc   c_1,t_1,c_1     !=
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    %g0,%g0,c_3
+        umul    a_3,b_6,t_1     !=!mul_add_c(a[3],b[6],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        umul    a_4,b_5,t_1     !mul_add_c(a[4],b[5],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    c_3,%g0,c_3
+        umul    a_5,b_4,t_1     !mul_add_c(a[5],b[4],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2          !=
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3
+        umul    a_6,b_3,t_1     !mul_add_c(a[6],b[3],c1,c2,c3);
+        addcc   c_1,t_1,c_1     !=
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3
+        umul    a_7,b_2,t_1     !=!mul_add_c(a[7],b[2],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        st      c_1,rp(9)       !r[9]=c1;
+
+        umul    a_7,b_3,t_1     !mul_add_c(a[7],b[3],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    %g0,%g0,c_1
+        umul    a_6,b_4,t_1     !mul_add_c(a[6],b[4],c2,c3,c1);
+        addcc   c_2,t_1,c_2     !=
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        umul    a_5,b_5,t_1     !=!mul_add_c(a[5],b[5],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1     !=
+        umul    a_4,b_6,t_1     !mul_add_c(a[4],b[6],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        umul    a_3,b_7,t_1     !mul_add_c(a[3],b[7],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        st      c_2,rp(10)      !r[10]=c2;
+
+        umul    a_4,b_7,t_1     !=!mul_add_c(a[4],b[7],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    %g0,%g0,c_2     !=
+        umul    a_5,b_6,t_1     !mul_add_c(a[5],b[6],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1     !=
+        addx    c_2,%g0,c_2
+        umul    a_6,b_5,t_1     !mul_add_c(a[6],b[5],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2          !=
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        umul    a_7,b_4,t_1     !mul_add_c(a[7],b[4],c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        st      c_3,rp(11)      !r[11]=c3;
+        addx    c_2,%g0,c_2     !=
+
+        umul    a_7,b_5,t_1     !mul_add_c(a[7],b[5],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    %g0,%g0,c_3
+        umul    a_6,b_6,t_1     !mul_add_c(a[6],b[6],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2          !=
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3
+        umul    a_5,b_7,t_1     !mul_add_c(a[5],b[7],c1,c2,c3);
+        addcc   c_1,t_1,c_1     !=
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        st      c_1,rp(12)      !r[12]=c1;
+        addx    c_3,%g0,c_3     !=
+
+        umul    a_6,b_7,t_1     !mul_add_c(a[6],b[7],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    %g0,%g0,c_1
+        umul    a_7,b_6,t_1     !mul_add_c(a[7],b[6],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        st      c_2,rp(13)      !r[13]=c2;
+
+        umul    a_7,b_7,t_1     !=!mul_add_c(a[7],b[7],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        nop                     !=
+        st      c_3,rp(14)      !r[14]=c3;
+        st      c_1,rp(15)      !r[15]=c1;
+
+        ret
+        restore %g0,%g0,%o0
+
+.type   bn_mul_comba8,#function
+.size   bn_mul_comba8,(.-bn_mul_comba8)
+
+.align  32
+
+.global bn_mul_comba4
+/*
+ * void bn_mul_comba4(r,a,b)
+ * BN_ULONG *r,*a,*b;
+ */
+bn_mul_comba4:
+        save    %sp,FRAME_SIZE,%sp
+        ld      ap(0),a_0
+        ld      bp(0),b_0
+        umul    a_0,b_0,c_1     !=!mul_add_c(a[0],b[0],c1,c2,c3);
+        ld      bp(1),b_1
+        rd      %y,c_2
+        st      c_1,rp(0)       !r[0]=c1;
+
+        umul    a_0,b_1,t_1     !=!mul_add_c(a[0],b[1],c2,c3,c1);
+        ld      ap(1),a_1
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  %g0,t_2,c_3
+        addx    %g0,%g0,c_1
+        ld      ap(2),a_2
+        umul    a_1,b_0,t_1     !=!mul_add_c(a[1],b[0],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1     !=
+        st      c_2,rp(1)       !r[1]=c2;
+
+        umul    a_2,b_0,t_1     !mul_add_c(a[2],b[0],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2          !=
+        addxcc  c_1,t_2,c_1
+        addx    %g0,%g0,c_2
+        ld      bp(2),b_2
+        umul    a_1,b_1,t_1     !=!mul_add_c(a[1],b[1],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2     !=
+        ld      bp(3),b_3
+        umul    a_0,b_2,t_1     !mul_add_c(a[0],b[2],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2          !=
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        st      c_3,rp(2)       !r[2]=c3;
+
+        umul    a_0,b_3,t_1     !=!mul_add_c(a[0],b[3],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    %g0,%g0,c_3     !=
+        umul    a_1,b_2,t_1     !mul_add_c(a[1],b[2],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    c_3,%g0,c_3
+        ld      ap(3),a_3
+        umul    a_2,b_1,t_1     !mul_add_c(a[2],b[1],c1,c2,c3);
+        addcc   c_1,t_1,c_1     !=
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3
+        umul    a_3,b_0,t_1     !=!mul_add_c(a[3],b[0],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        st      c_1,rp(3)       !r[3]=c1;
+
+        umul    a_3,b_1,t_1     !mul_add_c(a[3],b[1],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    %g0,%g0,c_1
+        umul    a_2,b_2,t_1     !mul_add_c(a[2],b[2],c2,c3,c1);
+        addcc   c_2,t_1,c_2     !=
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        umul    a_1,b_3,t_1     !=!mul_add_c(a[1],b[3],c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1     !=
+        st      c_2,rp(4)       !r[4]=c2;
+
+        umul    a_2,b_3,t_1     !mul_add_c(a[2],b[3],c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2          !=
+        addxcc  c_1,t_2,c_1
+        addx    %g0,%g0,c_2
+        umul    a_3,b_2,t_1     !mul_add_c(a[3],b[2],c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        st      c_3,rp(5)       !r[5]=c3;
+        addx    c_2,%g0,c_2     !=
+
+        umul    a_3,b_3,t_1     !mul_add_c(a[3],b[3],c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        st      c_1,rp(6)       !r[6]=c1;
+        st      c_2,rp(7)       !r[7]=c2;
+        
+        ret
+        restore %g0,%g0,%o0
+
+.type   bn_mul_comba4,#function
+.size   bn_mul_comba4,(.-bn_mul_comba4)
+
+.align  32
+
+.global bn_sqr_comba8
+bn_sqr_comba8:
+        save    %sp,FRAME_SIZE,%sp
+        ld      ap(0),a_0
+        ld      ap(1),a_1
+        umul    a_0,a_0,c_1     !=!sqr_add_c(a,0,c1,c2,c3);
+        rd      %y,c_2
+        st      c_1,rp(0)       !r[0]=c1;
+
+        ld      ap(2),a_2
+        umul    a_0,a_1,t_1     !=!sqr_add_c2(a,1,0,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  %g0,t_2,c_3
+        addx    %g0,%g0,c_1     !=
+        addcc   c_2,t_1,c_2
+        addxcc  c_3,t_2,c_3
+        st      c_2,rp(1)       !r[1]=c2;
+        addx    c_1,%g0,c_1     !=
+
+        umul    a_2,a_0,t_1     !sqr_add_c2(a,2,0,c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1     !=
+        addx    %g0,%g0,c_2
+        addcc   c_3,t_1,c_3
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2     !=
+        ld      ap(3),a_3
+        umul    a_1,a_1,t_1     !sqr_add_c(a,1,c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2          !=
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        st      c_3,rp(2)       !r[2]=c3;
+
+        umul    a_0,a_3,t_1     !=!sqr_add_c2(a,3,0,c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    %g0,%g0,c_3     !=
+        addcc   c_1,t_1,c_1
+        addxcc  c_2,t_2,c_2
+        ld      ap(4),a_4
+        addx    c_3,%g0,c_3     !=
+        umul    a_1,a_2,t_1     !sqr_add_c2(a,2,1,c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    c_3,%g0,c_3
+        addcc   c_1,t_1,c_1
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        st      c_1,rp(3)       !r[3]=c1;
+
+        umul    a_4,a_0,t_1     !sqr_add_c2(a,4,0,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    %g0,%g0,c_1
+        addcc   c_2,t_1,c_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        umul    a_3,a_1,t_1     !sqr_add_c2(a,3,1,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        addcc   c_2,t_1,c_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        ld      ap(5),a_5
+        umul    a_2,a_2,t_1     !sqr_add_c(a,2,c2,c3,c1);
+        addcc   c_2,t_1,c_2     !=
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        st      c_2,rp(4)       !r[4]=c2;
+        addx    c_1,%g0,c_1     !=
+
+        umul    a_0,a_5,t_1     !sqr_add_c2(a,5,0,c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1     !=
+        addx    %g0,%g0,c_2
+        addcc   c_3,t_1,c_3
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2     !=
+        umul    a_1,a_4,t_1     !sqr_add_c2(a,4,1,c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1     !=
+        addx    c_2,%g0,c_2
+        addcc   c_3,t_1,c_3
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2     !=
+        ld      ap(6),a_6
+        umul    a_2,a_3,t_1     !sqr_add_c2(a,3,2,c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2          !=
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        addcc   c_3,t_1,c_3
+        addxcc  c_1,t_2,c_1     !=
+        addx    c_2,%g0,c_2
+        st      c_3,rp(5)       !r[5]=c3;
+
+        umul    a_6,a_0,t_1     !sqr_add_c2(a,6,0,c1,c2,c3);
+        addcc   c_1,t_1,c_1     !=
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    %g0,%g0,c_3
+        addcc   c_1,t_1,c_1     !=
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3
+        umul    a_5,a_1,t_1     !sqr_add_c2(a,5,1,c1,c2,c3);
+        addcc   c_1,t_1,c_1     !=
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3
+        addcc   c_1,t_1,c_1     !=
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3
+        umul    a_4,a_2,t_1     !sqr_add_c2(a,4,2,c1,c2,c3);
+        addcc   c_1,t_1,c_1     !=
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3
+        addcc   c_1,t_1,c_1     !=
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3
+        ld      ap(7),a_7
+        umul    a_3,a_3,t_1     !=!sqr_add_c(a,3,c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        st      c_1,rp(6)       !r[6]=c1;
+
+        umul    a_0,a_7,t_1     !sqr_add_c2(a,7,0,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    %g0,%g0,c_1
+        addcc   c_2,t_1,c_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        umul    a_1,a_6,t_1     !sqr_add_c2(a,6,1,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        addcc   c_2,t_1,c_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        umul    a_2,a_5,t_1     !sqr_add_c2(a,5,2,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        addcc   c_2,t_1,c_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        umul    a_3,a_4,t_1     !sqr_add_c2(a,4,3,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        addcc   c_2,t_1,c_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        st      c_2,rp(7)       !r[7]=c2;
+
+        umul    a_7,a_1,t_1     !sqr_add_c2(a,7,1,c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    %g0,%g0,c_2
+        addcc   c_3,t_1,c_3     !=
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        umul    a_6,a_2,t_1     !sqr_add_c2(a,6,2,c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        addcc   c_3,t_1,c_3     !=
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        umul    a_5,a_3,t_1     !sqr_add_c2(a,5,3,c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        addcc   c_3,t_1,c_3     !=
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        umul    a_4,a_4,t_1     !sqr_add_c(a,4,c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        st      c_3,rp(8)       !r[8]=c3;
+        addx    c_2,%g0,c_2     !=
+
+        umul    a_2,a_7,t_1     !sqr_add_c2(a,7,2,c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    %g0,%g0,c_3
+        addcc   c_1,t_1,c_1
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        umul    a_3,a_6,t_1     !sqr_add_c2(a,6,3,c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    c_3,%g0,c_3
+        addcc   c_1,t_1,c_1
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        umul    a_4,a_5,t_1     !sqr_add_c2(a,5,4,c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    c_3,%g0,c_3
+        addcc   c_1,t_1,c_1
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        st      c_1,rp(9)       !r[9]=c1;
+
+        umul    a_7,a_3,t_1     !sqr_add_c2(a,7,3,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    %g0,%g0,c_1
+        addcc   c_2,t_1,c_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        umul    a_6,a_4,t_1     !sqr_add_c2(a,6,4,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        addcc   c_2,t_1,c_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        umul    a_5,a_5,t_1     !sqr_add_c(a,5,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        st      c_2,rp(10)      !r[10]=c2;
+
+        umul    a_4,a_7,t_1     !=!sqr_add_c2(a,7,4,c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    %g0,%g0,c_2     !=
+        addcc   c_3,t_1,c_3
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2
+        umul    a_5,a_6,t_1     !=!sqr_add_c2(a,6,5,c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    c_2,%g0,c_2     !=
+        addcc   c_3,t_1,c_3
+        addxcc  c_1,t_2,c_1
+        st      c_3,rp(11)      !r[11]=c3;
+        addx    c_2,%g0,c_2     !=
+
+        umul    a_7,a_5,t_1     !sqr_add_c2(a,7,5,c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    %g0,%g0,c_3
+        addcc   c_1,t_1,c_1
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        umul    a_6,a_6,t_1     !sqr_add_c(a,6,c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    c_3,%g0,c_3
+        st      c_1,rp(12)      !r[12]=c1;
+
+        umul    a_6,a_7,t_1     !sqr_add_c2(a,7,6,c2,c3,c1);
+        addcc   c_2,t_1,c_2     !=
+        rd      %y,t_2
+        addxcc  c_3,t_2,c_3
+        addx    %g0,%g0,c_1
+        addcc   c_2,t_1,c_2     !=
+        addxcc  c_3,t_2,c_3
+        st      c_2,rp(13)      !r[13]=c2;
+        addx    c_1,%g0,c_1     !=
+
+        umul    a_7,a_7,t_1     !sqr_add_c(a,7,c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1     !=
+        st      c_3,rp(14)      !r[14]=c3;
+        st      c_1,rp(15)      !r[15]=c1;
+
+        ret
+        restore %g0,%g0,%o0
+
+.type   bn_sqr_comba8,#function
+.size   bn_sqr_comba8,(.-bn_sqr_comba8)
+
+.align  32
+
+.global bn_sqr_comba4
+/*
+ * void bn_sqr_comba4(r,a)
+ * BN_ULONG *r,*a;
+ */
+bn_sqr_comba4:
+        save    %sp,FRAME_SIZE,%sp
+        ld      ap(0),a_0
+        umul    a_0,a_0,c_1     !sqr_add_c(a,0,c1,c2,c3);
+        ld      ap(1),a_1       !=
+        rd      %y,c_2
+        st      c_1,rp(0)       !r[0]=c1;
+
+        ld      ap(2),a_2
+        umul    a_0,a_1,t_1     !=!sqr_add_c2(a,1,0,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2
+        addxcc  %g0,t_2,c_3
+        addx    %g0,%g0,c_1     !=
+        addcc   c_2,t_1,c_2
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1     !=
+        st      c_2,rp(1)       !r[1]=c2;
+
+        umul    a_2,a_0,t_1     !sqr_add_c2(a,2,0,c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2          !=
+        addxcc  c_1,t_2,c_1
+        addx    %g0,%g0,c_2
+        addcc   c_3,t_1,c_3
+        addxcc  c_1,t_2,c_1     !=
+        addx    c_2,%g0,c_2
+        ld      ap(3),a_3
+        umul    a_1,a_1,t_1     !sqr_add_c(a,1,c3,c1,c2);
+        addcc   c_3,t_1,c_3     !=
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        st      c_3,rp(2)       !r[2]=c3;
+        addx    c_2,%g0,c_2     !=
+
+        umul    a_0,a_3,t_1     !sqr_add_c2(a,3,0,c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    %g0,%g0,c_3
+        addcc   c_1,t_1,c_1
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        umul    a_1,a_2,t_1     !sqr_add_c2(a,2,1,c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        addx    c_3,%g0,c_3
+        addcc   c_1,t_1,c_1
+        addxcc  c_2,t_2,c_2
+        addx    c_3,%g0,c_3     !=
+        st      c_1,rp(3)       !r[3]=c1;
+
+        umul    a_3,a_1,t_1     !sqr_add_c2(a,3,1,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    %g0,%g0,c_1
+        addcc   c_2,t_1,c_2
+        addxcc  c_3,t_2,c_3     !=
+        addx    c_1,%g0,c_1
+        umul    a_2,a_2,t_1     !sqr_add_c(a,2,c2,c3,c1);
+        addcc   c_2,t_1,c_2
+        rd      %y,t_2          !=
+        addxcc  c_3,t_2,c_3
+        addx    c_1,%g0,c_1
+        st      c_2,rp(4)       !r[4]=c2;
+
+        umul    a_2,a_3,t_1     !=!sqr_add_c2(a,3,2,c3,c1,c2);
+        addcc   c_3,t_1,c_3
+        rd      %y,t_2
+        addxcc  c_1,t_2,c_1
+        addx    %g0,%g0,c_2     !=
+        addcc   c_3,t_1,c_3
+        addxcc  c_1,t_2,c_1
+        st      c_3,rp(5)       !r[5]=c3;
+        addx    c_2,%g0,c_2     !=
+
+        umul    a_3,a_3,t_1     !sqr_add_c(a,3,c1,c2,c3);
+        addcc   c_1,t_1,c_1
+        rd      %y,t_2
+        addxcc  c_2,t_2,c_2     !=
+        st      c_1,rp(6)       !r[6]=c1;
+        st      c_2,rp(7)       !r[7]=c2;
+        
+        ret
+        restore %g0,%g0,%o0
+
+.type   bn_sqr_comba4,#function
+.size   bn_sqr_comba4,(.-bn_sqr_comba4)
+
+.align  32
diff --git a/src/lib/libcrypto/bn/asm/sparcv8plus.S b/src/lib/libcrypto/bn/asm/sparcv8plus.S
new file mode 100644
index 0000000000..0074dfdb75
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/sparcv8plus.S
@@ -0,0 +1,1535 @@
+.ident  "sparcv8plus.s, Version 1.4"
+.ident  "SPARC v9 ISA artwork by Andy Polyakov <appro@fy.chalmers.se>"
+
+/*
+ * ====================================================================
+ * Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+ * project.
+ *
+ * Rights for redistribution and usage in source and binary forms are
+ * granted according to the OpenSSL license. Warranty of any kind is
+ * disclaimed.
+ * ====================================================================
+ */
+
+/*
+ * This is my modest contributon to OpenSSL project (see
+ * http://www.openssl.org/ for more information about it) and is
+ * a drop-in UltraSPARC ISA replacement for crypto/bn/bn_asm.c
+ * module. For updates see http://fy.chalmers.se/~appro/hpe/.
+ *
+ * Questions-n-answers.
+ *
+ * Q. How to compile?
+ * A. With SC4.x/SC5.x:
+ *
+ *      cc -xarch=v8plus -c bn_asm.sparc.v8plus.S -o bn_asm.o
+ *
+ *    and with gcc:
+ *
+ *      gcc -mcpu=ultrasparc -c bn_asm.sparc.v8plus.S -o bn_asm.o
+ *
+ *    or if above fails (it does if you have gas installed):
+ *
+ *      gcc -E bn_asm.sparc.v8plus.S | as -xarch=v8plus /dev/fd/0 -o bn_asm.o
+ *
+ *    Quick-n-dirty way to fuse the module into the library.
+ *    Provided that the library is already configured and built
+ *    (in 0.9.2 case with no-asm option):
+ *
+ *      # cd crypto/bn
+ *      # cp /some/place/bn_asm.sparc.v8plus.S .
+ *      # cc -xarch=v8plus -c bn_asm.sparc.v8plus.S -o bn_asm.o
+ *      # make
+ *      # cd ../..
+ *      # make; make test
+ *
+ *    Quick-n-dirty way to get rid of it:
+ *
+ *      # cd crypto/bn
+ *      # touch bn_asm.c
+ *      # make
+ *      # cd ../..
+ *      # make; make test
+ *
+ * Q. V8plus achitecture? What kind of beast is that?
+ * A. Well, it's rather a programming model than an architecture...
+ *    It's actually v9-compliant, i.e. *any* UltraSPARC, CPU under
+ *    special conditions, namely when kernel doesn't preserve upper
+ *    32 bits of otherwise 64-bit registers during a context switch.
+ *
+ * Q. Why just UltraSPARC? What about SuperSPARC?
+ * A. Original release did target UltraSPARC only. Now SuperSPARC
+ *    version is provided along. Both version share bn_*comba[48]
+ *    implementations (see comment later in code for explanation).
+ *    But what's so special about this UltraSPARC implementation?
+ *    Why didn't I let compiler do the job? Trouble is that most of
+ *    available compilers (well, SC5.0 is the only exception) don't
+ *    attempt to take advantage of UltraSPARC's 64-bitness under
+ *    32-bit kernels even though it's perfectly possible (see next
+ *    question).
+ *
+ * Q. 64-bit registers under 32-bit kernels? Didn't you just say it
+ *    doesn't work?
+ * A. You can't adress *all* registers as 64-bit wide:-( The catch is
+ *    that you actually may rely upon %o0-%o5 and %g1-%g4 being fully
+ *    preserved if you're in a leaf function, i.e. such never calling
+ *    any other functions. All functions in this module are leaf and
+ *    10 registers is a handful. And as a matter of fact none-"comba"
+ *    routines don't require even that much and I could even afford to
+ *    not allocate own stack frame for 'em:-)
+ *
+ * Q. What about 64-bit kernels?
+ * A. What about 'em? Just kidding:-) Pure 64-bit version is currently
+ *    under evaluation and development...
+ *
+ * Q. What about shared libraries?
+ * A. What about 'em? Kidding again:-) Code does *not* contain any
+ *    code position dependencies and it's safe to include it into
+ *    shared library as is.
+ *
+ * Q. How much faster does it go?
+ * A. Do you have a good benchmark? In either case below is what I
+ *    experience with crypto/bn/expspeed.c test program:
+ *
+ *      v8plus module on U10/300MHz against bn_asm.c compiled with:
+ *
+ *      cc-5.0 -xarch=v8plus -xO5 -xdepend      +7-12%
+ *      cc-4.2 -xarch=v8plus -xO5 -xdepend      +25-35%
+ *      egcs-1.1.2 -mcpu=ultrasparc -O3         +35-45%
+ *
+ *      v8 module on SS10/60MHz against bn_asm.c compiled with:
+ *
+ *      cc-5.0 -xarch=v8 -xO5 -xdepend          +7-10%
+ *      cc-4.2 -xarch=v8 -xO5 -xdepend          +10%
+ *      egcs-1.1.2 -mv8 -O3                     +35-45%
+ *
+ *    As you can see it's damn hard to beat the new Sun C compiler
+ *    and it's in first place GNU C users who will appreciate this
+ *    assembler implementation:-)       
+ */
+
+/*
+ * Revision history.
+ *
+ * 1.0  - initial release;
+ * 1.1  - new loop unrolling model(*);
+ *      - some more fine tuning;
+ * 1.2  - made gas friendly;
+ *      - updates to documentation concerning v9;
+ *      - new performance comparison matrix;
+ * 1.3  - fixed problem with /usr/ccs/lib/cpp;
+ * 1.4  - native V9 bn_*_comba[48] implementation (15% more efficient)
+ *        resulting in slight overall performance kick;
+ *      - some retunes;
+ *      - support for GNU as added;
+ *
+ * (*)  Originally unrolled loop looked like this:
+ *          for (;;) {
+ *              op(p+0); if (--n==0) break;
+ *              op(p+1); if (--n==0) break;
+ *              op(p+2); if (--n==0) break;
+ *              op(p+3); if (--n==0) break;
+ *              p+=4;
+ *          }
+ *      I unroll according to following:
+ *          while (n&~3) {
+ *              op(p+0); op(p+1); op(p+2); op(p+3);
+ *              p+=4; n=-4;
+ *          }
+ *          if (n) {
+ *              op(p+0); if (--n==0) return;
+ *              op(p+2); if (--n==0) return;
+ *              op(p+3); return;
+ *          }
+ */
+
+/*
+ * GNU assembler can't stand stuw:-(
+ */
+#define stuw st
+
+.section        ".text",#alloc,#execinstr
+.file           "bn_asm.sparc.v8plus.S"
+
+.align  32
+
+.global bn_mul_add_words
+/*
+ * BN_ULONG bn_mul_add_words(rp,ap,num,w)
+ * BN_ULONG *rp,*ap;
+ * int num;
+ * BN_ULONG w;
+ */
+bn_mul_add_words:
+        brgz,a  %o2,.L_bn_mul_add_words_proceed
+        lduw    [%o1],%g2
+        retl
+        clr     %o0
+
+.L_bn_mul_add_words_proceed:
+        srl     %o3,%g0,%o3     ! clruw %o3
+        andcc   %o2,-4,%g0
+        bz,pn   %icc,.L_bn_mul_add_words_tail
+        clr     %o5
+
+.L_bn_mul_add_words_loop:       ! wow! 32 aligned!
+        lduw    [%o0],%g1
+        lduw    [%o1+4],%g3
+        mulx    %o3,%g2,%g2
+        add     %g1,%o5,%o4
+        nop
+        add     %o4,%g2,%o4
+        stuw    %o4,[%o0]
+        srlx    %o4,32,%o5
+
+        lduw    [%o0+4],%g1
+        lduw    [%o1+8],%g2
+        mulx    %o3,%g3,%g3
+        add     %g1,%o5,%o4
+        dec     4,%o2
+        add     %o4,%g3,%o4
+        stuw    %o4,[%o0+4]
+        srlx    %o4,32,%o5
+
+        lduw    [%o0+8],%g1
+        lduw    [%o1+12],%g3
+        mulx    %o3,%g2,%g2
+        add     %g1,%o5,%o4
+        inc     16,%o1
+        add     %o4,%g2,%o4
+        stuw    %o4,[%o0+8]
+        srlx    %o4,32,%o5
+
+        lduw    [%o0+12],%g1
+        mulx    %o3,%g3,%g3
+        add     %g1,%o5,%o4
+        inc     16,%o0
+        add     %o4,%g3,%o4
+        andcc   %o2,-4,%g0
+        stuw    %o4,[%o0-4]
+        srlx    %o4,32,%o5
+        bnz,a,pt        %icc,.L_bn_mul_add_words_loop
+        lduw    [%o1],%g2
+
+        brnz,a,pn       %o2,.L_bn_mul_add_words_tail
+        lduw    [%o1],%g2
+.L_bn_mul_add_words_return:
+        retl
+        mov     %o5,%o0
+
+.L_bn_mul_add_words_tail:
+        lduw    [%o0],%g1
+        mulx    %o3,%g2,%g2
+        add     %g1,%o5,%o4
+        dec     %o2
+        add     %o4,%g2,%o4
+        srlx    %o4,32,%o5
+        brz,pt  %o2,.L_bn_mul_add_words_return
+        stuw    %o4,[%o0]
+
+        lduw    [%o1+4],%g2
+        lduw    [%o0+4],%g1
+        mulx    %o3,%g2,%g2
+        add     %g1,%o5,%o4
+        dec     %o2
+        add     %o4,%g2,%o4
+        srlx    %o4,32,%o5
+        brz,pt  %o2,.L_bn_mul_add_words_return
+        stuw    %o4,[%o0+4]
+
+        lduw    [%o1+8],%g2
+        lduw    [%o0+8],%g1
+        mulx    %o3,%g2,%g2
+        add     %g1,%o5,%o4
+        add     %o4,%g2,%o4
+        stuw    %o4,[%o0+8]
+        retl
+        srlx    %o4,32,%o0
+
+.type   bn_mul_add_words,#function
+.size   bn_mul_add_words,(.-bn_mul_add_words)
+
+.align  32
+
+.global bn_mul_words
+/*
+ * BN_ULONG bn_mul_words(rp,ap,num,w)
+ * BN_ULONG *rp,*ap;
+ * int num;
+ * BN_ULONG w;
+ */
+bn_mul_words:
+        brgz,a  %o2,.L_bn_mul_words_proceeed
+        lduw    [%o1],%g2
+        retl
+        clr     %o0
+
+.L_bn_mul_words_proceeed:
+        srl     %o3,%g0,%o3     ! clruw %o3
+        andcc   %o2,-4,%g0
+        bz,pn   %icc,.L_bn_mul_words_tail
+        clr     %o5
+
+.L_bn_mul_words_loop:           ! wow! 32 aligned!
+        lduw    [%o1+4],%g3
+        mulx    %o3,%g2,%g2
+        add     %g2,%o5,%o4
+        nop
+        stuw    %o4,[%o0]
+        srlx    %o4,32,%o5
+
+        lduw    [%o1+8],%g2
+        mulx    %o3,%g3,%g3
+        add     %g3,%o5,%o4
+        dec     4,%o2
+        stuw    %o4,[%o0+4]
+        srlx    %o4,32,%o5
+
+        lduw    [%o1+12],%g3
+        mulx    %o3,%g2,%g2
+        add     %g2,%o5,%o4
+        inc     16,%o1
+        stuw    %o4,[%o0+8]
+        srlx    %o4,32,%o5
+
+        mulx    %o3,%g3,%g3
+        add     %g3,%o5,%o4
+        inc     16,%o0
+        stuw    %o4,[%o0-4]
+        srlx    %o4,32,%o5
+        andcc   %o2,-4,%g0
+        bnz,a,pt        %icc,.L_bn_mul_words_loop
+        lduw    [%o1],%g2
+        nop
+        nop
+
+        brnz,a,pn       %o2,.L_bn_mul_words_tail
+        lduw    [%o1],%g2
+.L_bn_mul_words_return:
+        retl
+        mov     %o5,%o0
+
+.L_bn_mul_words_tail:
+        mulx    %o3,%g2,%g2
+        add     %g2,%o5,%o4
+        dec     %o2
+        srlx    %o4,32,%o5
+        brz,pt  %o2,.L_bn_mul_words_return
+        stuw    %o4,[%o0]
+
+        lduw    [%o1+4],%g2
+        mulx    %o3,%g2,%g2
+        add     %g2,%o5,%o4
+        dec     %o2
+        srlx    %o4,32,%o5
+        brz,pt  %o2,.L_bn_mul_words_return
+        stuw    %o4,[%o0+4]
+
+        lduw    [%o1+8],%g2
+        mulx    %o3,%g2,%g2
+        add     %g2,%o5,%o4
+        stuw    %o4,[%o0+8]
+        retl
+        srlx    %o4,32,%o0
+
+.type   bn_mul_words,#function
+.size   bn_mul_words,(.-bn_mul_words)
+
+.align  32
+.global bn_sqr_words
+/*
+ * void bn_sqr_words(r,a,n)
+ * BN_ULONG *r,*a;
+ * int n;
+ */
+bn_sqr_words:
+        brgz,a  %o2,.L_bn_sqr_words_proceeed
+        lduw    [%o1],%g2
+        retl
+        clr     %o0
+
+.L_bn_sqr_words_proceeed:
+        andcc   %o2,-4,%g0
+        nop
+        bz,pn   %icc,.L_bn_sqr_words_tail
+        nop
+
+.L_bn_sqr_words_loop:           ! wow! 32 aligned!
+        lduw    [%o1+4],%g3
+        mulx    %g2,%g2,%o4
+        stuw    %o4,[%o0]
+        srlx    %o4,32,%o5
+        stuw    %o5,[%o0+4]
+        nop
+
+        lduw    [%o1+8],%g2
+        mulx    %g3,%g3,%o4
+        dec     4,%o2
+        stuw    %o4,[%o0+8]
+        srlx    %o4,32,%o5
+        stuw    %o5,[%o0+12]
+
+        lduw    [%o1+12],%g3
+        mulx    %g2,%g2,%o4
+        srlx    %o4,32,%o5
+        stuw    %o4,[%o0+16]
+        inc     16,%o1
+        stuw    %o5,[%o0+20]
+
+        mulx    %g3,%g3,%o4
+        inc     32,%o0
+        stuw    %o4,[%o0-8]
+        srlx    %o4,32,%o5
+        andcc   %o2,-4,%g2
+        stuw    %o5,[%o0-4]
+        bnz,a,pt        %icc,.L_bn_sqr_words_loop
+        lduw    [%o1],%g2
+        nop
+
+        brnz,a,pn       %o2,.L_bn_sqr_words_tail
+        lduw    [%o1],%g2
+.L_bn_sqr_words_return:
+        retl
+        clr     %o0
+
+.L_bn_sqr_words_tail:
+        mulx    %g2,%g2,%o4
+        dec     %o2
+        stuw    %o4,[%o0]
+        srlx    %o4,32,%o5
+        brz,pt  %o2,.L_bn_sqr_words_return
+        stuw    %o5,[%o0+4]
+
+        lduw    [%o1+4],%g2
+        mulx    %g2,%g2,%o4
+        dec     %o2
+        stuw    %o4,[%o0+8]
+        srlx    %o4,32,%o5
+        brz,pt  %o2,.L_bn_sqr_words_return
+        stuw    %o5,[%o0+12]
+
+        lduw    [%o1+8],%g2
+        mulx    %g2,%g2,%o4
+        srlx    %o4,32,%o5
+        stuw    %o4,[%o0+16]
+        stuw    %o5,[%o0+20]
+        retl
+        clr     %o0
+
+.type   bn_sqr_words,#function
+.size   bn_sqr_words,(.-bn_sqr_words)
+
+.align  32
+.global bn_div_words
+/*
+ * BN_ULONG bn_div_words(h,l,d)
+ * BN_ULONG h,l,d;
+ */
+bn_div_words:
+        sllx    %o0,32,%o0
+        or      %o0,%o1,%o0
+        udivx   %o0,%o2,%o0
+        retl
+        srl     %o0,%g0,%o0     ! clruw %o0
+
+.type   bn_div_words,#function
+.size   bn_div_words,(.-bn_div_words)
+
+.align  32
+
+.global bn_add_words
+/*
+ * BN_ULONG bn_add_words(rp,ap,bp,n)
+ * BN_ULONG *rp,*ap,*bp;
+ * int n;
+ */
+bn_add_words:
+        brgz,a  %o3,.L_bn_add_words_proceed
+        lduw    [%o1],%o4
+        retl
+        clr     %o0
+
+.L_bn_add_words_proceed:
+        andcc   %o3,-4,%g0
+        bz,pn   %icc,.L_bn_add_words_tail
+        addcc   %g0,0,%g0       ! clear carry flag
+        nop
+
+.L_bn_add_words_loop:           ! wow! 32 aligned!
+        dec     4,%o3
+        lduw    [%o2],%o5
+        lduw    [%o1+4],%g1
+        lduw    [%o2+4],%g2
+        lduw    [%o1+8],%g3
+        lduw    [%o2+8],%g4
+        addccc  %o5,%o4,%o5
+        stuw    %o5,[%o0]
+
+        lduw    [%o1+12],%o4
+        lduw    [%o2+12],%o5
+        inc     16,%o1
+        addccc  %g1,%g2,%g1
+        stuw    %g1,[%o0+4]
+        
+        inc     16,%o2
+        addccc  %g3,%g4,%g3
+        stuw    %g3,[%o0+8]
+
+        inc     16,%o0
+        addccc  %o5,%o4,%o5
+        stuw    %o5,[%o0-4]
+        and     %o3,-4,%g1
+        brnz,a,pt       %g1,.L_bn_add_words_loop
+        lduw    [%o1],%o4
+
+        brnz,a,pn       %o3,.L_bn_add_words_tail
+        lduw    [%o1],%o4
+.L_bn_add_words_return:
+        clr     %o0
+        retl
+        movcs   %icc,1,%o0
+        nop
+
+.L_bn_add_words_tail:
+        lduw    [%o2],%o5
+        dec     %o3
+        addccc  %o5,%o4,%o5
+        brz,pt  %o3,.L_bn_add_words_return
+        stuw    %o5,[%o0]
+
+        lduw    [%o1+4],%o4
+        lduw    [%o2+4],%o5
+        dec     %o3
+        addccc  %o5,%o4,%o5
+        brz,pt  %o3,.L_bn_add_words_return
+        stuw    %o5,[%o0+4]
+
+        lduw    [%o1+8],%o4
+        lduw    [%o2+8],%o5
+        addccc  %o5,%o4,%o5
+        stuw    %o5,[%o0+8]
+        clr     %o0
+        retl
+        movcs   %icc,1,%o0
+
+.type   bn_add_words,#function
+.size   bn_add_words,(.-bn_add_words)
+
+.global bn_sub_words
+/*
+ * BN_ULONG bn_sub_words(rp,ap,bp,n)
+ * BN_ULONG *rp,*ap,*bp;
+ * int n;
+ */
+bn_sub_words:
+        brgz,a  %o3,.L_bn_sub_words_proceed
+        lduw    [%o1],%o4
+        retl
+        clr     %o0
+
+.L_bn_sub_words_proceed:
+        andcc   %o3,-4,%g0
+        bz,pn   %icc,.L_bn_sub_words_tail
+        addcc   %g0,0,%g0       ! clear carry flag
+        nop
+
+.L_bn_sub_words_loop:           ! wow! 32 aligned!
+        dec     4,%o3
+        lduw    [%o2],%o5
+        lduw    [%o1+4],%g1
+        lduw    [%o2+4],%g2
+        lduw    [%o1+8],%g3
+        lduw    [%o2+8],%g4
+        subccc  %o4,%o5,%o5
+        stuw    %o5,[%o0]
+
+        lduw    [%o1+12],%o4
+        lduw    [%o2+12],%o5
+        inc     16,%o1
+        subccc  %g1,%g2,%g2
+        stuw    %g2,[%o0+4]
+
+        inc     16,%o2
+        subccc  %g3,%g4,%g4
+        stuw    %g4,[%o0+8]
+
+        inc     16,%o0
+        subccc  %o4,%o5,%o5
+        stuw    %o5,[%o0-4]
+        and     %o3,-4,%g1
+        brnz,a,pt       %g1,.L_bn_sub_words_loop
+        lduw    [%o1],%o4
+
+        brnz,a,pn       %o3,.L_bn_sub_words_tail
+        lduw    [%o1],%o4
+.L_bn_sub_words_return:
+        clr     %o0
+        retl
+        movcs   %icc,1,%o0
+        nop
+
+.L_bn_sub_words_tail:           ! wow! 32 aligned!
+        lduw    [%o2],%o5
+        dec     %o3
+        subccc  %o4,%o5,%o5
+        brz,pt  %o3,.L_bn_sub_words_return
+        stuw    %o5,[%o0]
+
+        lduw    [%o1+4],%o4
+        lduw    [%o2+4],%o5
+        dec     %o3
+        subccc  %o4,%o5,%o5
+        brz,pt  %o3,.L_bn_sub_words_return
+        stuw    %o5,[%o0+4]
+
+        lduw    [%o1+8],%o4
+        lduw    [%o2+8],%o5
+        subccc  %o4,%o5,%o5
+        stuw    %o5,[%o0+8]
+        clr     %o0
+        retl
+        movcs   %icc,1,%o0
+
+.type   bn_sub_words,#function
+.size   bn_sub_words,(.-bn_sub_words)
+
+/*
+ * Code below depends on the fact that upper parts of the %l0-%l7
+ * and %i0-%i7 are zeroed by kernel after context switch. In
+ * previous versions this comment stated that "the trouble is that
+ * it's not feasible to implement the mumbo-jumbo in less V9
+ * instructions:-(" which apparently isn't true thanks to
+ * 'bcs,a %xcc,.+8; inc %rd' pair. But the performance improvement
+ * results not from the shorter code, but from elimination of
+ * multicycle none-pairable 'rd %y,%rd' instructions.
+ *
+ *                                                      Andy.
+ */
+
+#define FRAME_SIZE      -96
+
+/*
+ * Here is register usage map for *all* routines below.
+ */
+#define t_1     %o0
+#define t_2     %o1
+#define c_12    %o2
+#define c_3     %o3
+
+#define ap(I)   [%i1+4*I]
+#define bp(I)   [%i2+4*I]
+#define rp(I)   [%i0+4*I]
+
+#define a_0     %l0
+#define a_1     %l1
+#define a_2     %l2
+#define a_3     %l3
+#define a_4     %l4
+#define a_5     %l5
+#define a_6     %l6
+#define a_7     %l7
+
+#define b_0     %i3
+#define b_1     %i4
+#define b_2     %i5
+#define b_3     %o4
+#define b_4     %o5
+#define b_5     %o7
+#define b_6     %g1
+#define b_7     %g4
+
+.align  32
+.global bn_mul_comba8
+/*
+ * void bn_mul_comba8(r,a,b)
+ * BN_ULONG *r,*a,*b;
+ */
+bn_mul_comba8:
+        save    %sp,FRAME_SIZE,%sp
+        mov     1,t_2
+        lduw    ap(0),a_0
+        sllx    t_2,32,t_2
+        lduw    bp(0),b_0       !=
+        lduw    bp(1),b_1
+        mulx    a_0,b_0,t_1     !mul_add_c(a[0],b[0],c1,c2,c3);
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(0)       !=!r[0]=c1;
+
+        lduw    ap(1),a_1
+        mulx    a_0,b_1,t_1     !mul_add_c(a[0],b[1],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3             !=
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    ap(2),a_2
+        mulx    a_1,b_0,t_1     !=!mul_add_c(a[1],b[0],c2,c3,c1);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12     !=
+        stuw    t_1,rp(1)       !r[1]=c2;
+        or      c_12,c_3,c_12
+
+        mulx    a_2,b_0,t_1     !mul_add_c(a[2],b[0],c3,c1,c2);
+        addcc   c_12,t_1,c_12   !=
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    bp(2),b_2       !=
+        mulx    a_1,b_1,t_1     !mul_add_c(a[1],b[1],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        lduw    bp(3),b_3
+        mulx    a_0,b_2,t_1     !mul_add_c(a[0],b[2],c3,c1,c2);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(2)       !r[2]=c3;
+        or      c_12,c_3,c_12   !=
+
+        mulx    a_0,b_3,t_1     !mul_add_c(a[0],b[3],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_1,b_2,t_1     !=!mul_add_c(a[1],b[2],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        lduw    ap(3),a_3
+        mulx    a_2,b_1,t_1     !mul_add_c(a[2],b[1],c1,c2,c3);
+        addcc   c_12,t_1,c_12   !=
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    ap(4),a_4
+        mulx    a_3,b_0,t_1     !=!mul_add_c(a[3],b[0],c1,c2,c3);!=
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12     !=
+        stuw    t_1,rp(3)       !r[3]=c1;
+        or      c_12,c_3,c_12
+
+        mulx    a_4,b_0,t_1     !mul_add_c(a[4],b[0],c2,c3,c1);
+        addcc   c_12,t_1,c_12   !=
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_3,b_1,t_1     !=!mul_add_c(a[3],b[1],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_2,b_2,t_1     !=!mul_add_c(a[2],b[2],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    bp(4),b_4       !=
+        mulx    a_1,b_3,t_1     !mul_add_c(a[1],b[3],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        lduw    bp(5),b_5
+        mulx    a_0,b_4,t_1     !mul_add_c(a[0],b[4],c2,c3,c1);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(4)       !r[4]=c2;
+        or      c_12,c_3,c_12   !=
+
+        mulx    a_0,b_5,t_1     !mul_add_c(a[0],b[5],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_1,b_4,t_1     !mul_add_c(a[1],b[4],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_2,b_3,t_1     !mul_add_c(a[2],b[3],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_3,b_2,t_1     !mul_add_c(a[3],b[2],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        lduw    ap(5),a_5
+        mulx    a_4,b_1,t_1     !mul_add_c(a[4],b[1],c3,c1,c2);
+        addcc   c_12,t_1,c_12   !=
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    ap(6),a_6
+        mulx    a_5,b_0,t_1     !=!mul_add_c(a[5],b[0],c3,c1,c2);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12     !=
+        stuw    t_1,rp(5)       !r[5]=c3;
+        or      c_12,c_3,c_12
+
+        mulx    a_6,b_0,t_1     !mul_add_c(a[6],b[0],c1,c2,c3);
+        addcc   c_12,t_1,c_12   !=
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_5,b_1,t_1     !=!mul_add_c(a[5],b[1],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_4,b_2,t_1     !=!mul_add_c(a[4],b[2],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_3,b_3,t_1     !=!mul_add_c(a[3],b[3],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_2,b_4,t_1     !=!mul_add_c(a[2],b[4],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    bp(6),b_6       !=
+        mulx    a_1,b_5,t_1     !mul_add_c(a[1],b[5],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        lduw    bp(7),b_7
+        mulx    a_0,b_6,t_1     !mul_add_c(a[0],b[6],c1,c2,c3);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(6)       !r[6]=c1;
+        or      c_12,c_3,c_12   !=
+
+        mulx    a_0,b_7,t_1     !mul_add_c(a[0],b[7],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_1,b_6,t_1     !mul_add_c(a[1],b[6],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_2,b_5,t_1     !mul_add_c(a[2],b[5],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_3,b_4,t_1     !mul_add_c(a[3],b[4],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_4,b_3,t_1     !mul_add_c(a[4],b[3],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_5,b_2,t_1     !mul_add_c(a[5],b[2],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        lduw    ap(7),a_7
+        mulx    a_6,b_1,t_1     !=!mul_add_c(a[6],b[1],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_7,b_0,t_1     !=!mul_add_c(a[7],b[0],c2,c3,c1);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12     !=
+        stuw    t_1,rp(7)       !r[7]=c2;
+        or      c_12,c_3,c_12
+
+        mulx    a_7,b_1,t_1     !=!mul_add_c(a[7],b[1],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        mulx    a_6,b_2,t_1     !mul_add_c(a[6],b[2],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        mulx    a_5,b_3,t_1     !mul_add_c(a[5],b[3],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        mulx    a_4,b_4,t_1     !mul_add_c(a[4],b[4],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        mulx    a_3,b_5,t_1     !mul_add_c(a[3],b[5],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        mulx    a_2,b_6,t_1     !mul_add_c(a[2],b[6],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        mulx    a_1,b_7,t_1     !mul_add_c(a[1],b[7],c3,c1,c2);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(8)       !r[8]=c3;
+        or      c_12,c_3,c_12
+
+        mulx    a_2,b_7,t_1     !=!mul_add_c(a[2],b[7],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        mulx    a_3,b_6,t_1     !mul_add_c(a[3],b[6],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_4,b_5,t_1     !mul_add_c(a[4],b[5],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_5,b_4,t_1     !mul_add_c(a[5],b[4],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_6,b_3,t_1     !mul_add_c(a[6],b[3],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_7,b_2,t_1     !mul_add_c(a[7],b[2],c1,c2,c3);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(9)       !r[9]=c1;
+        or      c_12,c_3,c_12   !=
+
+        mulx    a_7,b_3,t_1     !mul_add_c(a[7],b[3],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_6,b_4,t_1     !mul_add_c(a[6],b[4],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_5,b_5,t_1     !mul_add_c(a[5],b[5],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_4,b_6,t_1     !mul_add_c(a[4],b[6],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_3,b_7,t_1     !mul_add_c(a[3],b[7],c2,c3,c1);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(10)      !r[10]=c2;
+        or      c_12,c_3,c_12   !=
+
+        mulx    a_4,b_7,t_1     !mul_add_c(a[4],b[7],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_5,b_6,t_1     !mul_add_c(a[5],b[6],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_6,b_5,t_1     !mul_add_c(a[6],b[5],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_7,b_4,t_1     !mul_add_c(a[7],b[4],c3,c1,c2);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(11)      !r[11]=c3;
+        or      c_12,c_3,c_12   !=
+
+        mulx    a_7,b_5,t_1     !mul_add_c(a[7],b[5],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_6,b_6,t_1     !mul_add_c(a[6],b[6],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_5,b_7,t_1     !mul_add_c(a[5],b[7],c1,c2,c3);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(12)      !r[12]=c1;
+        or      c_12,c_3,c_12   !=
+
+        mulx    a_6,b_7,t_1     !mul_add_c(a[6],b[7],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_7,b_6,t_1     !mul_add_c(a[7],b[6],c2,c3,c1);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        st      t_1,rp(13)      !r[13]=c2;
+        or      c_12,c_3,c_12   !=
+
+        mulx    a_7,b_7,t_1     !mul_add_c(a[7],b[7],c3,c1,c2);
+        addcc   c_12,t_1,t_1
+        srlx    t_1,32,c_12     !=
+        stuw    t_1,rp(14)      !r[14]=c3;
+        stuw    c_12,rp(15)     !r[15]=c1;
+
+        ret
+        restore %g0,%g0,%o0     !=
+
+.type   bn_mul_comba8,#function
+.size   bn_mul_comba8,(.-bn_mul_comba8)
+
+.align  32
+
+.global bn_mul_comba4
+/*
+ * void bn_mul_comba4(r,a,b)
+ * BN_ULONG *r,*a,*b;
+ */
+bn_mul_comba4:
+        save    %sp,FRAME_SIZE,%sp
+        lduw    ap(0),a_0
+        mov     1,t_2
+        lduw    bp(0),b_0
+        sllx    t_2,32,t_2      !=
+        lduw    bp(1),b_1
+        mulx    a_0,b_0,t_1     !mul_add_c(a[0],b[0],c1,c2,c3);
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(0)       !=!r[0]=c1;
+
+        lduw    ap(1),a_1
+        mulx    a_0,b_1,t_1     !mul_add_c(a[0],b[1],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3             !=
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    ap(2),a_2
+        mulx    a_1,b_0,t_1     !=!mul_add_c(a[1],b[0],c2,c3,c1);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12     !=
+        stuw    t_1,rp(1)       !r[1]=c2;
+        or      c_12,c_3,c_12
+
+        mulx    a_2,b_0,t_1     !mul_add_c(a[2],b[0],c3,c1,c2);
+        addcc   c_12,t_1,c_12   !=
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    bp(2),b_2       !=
+        mulx    a_1,b_1,t_1     !mul_add_c(a[1],b[1],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3     !=
+        lduw    bp(3),b_3
+        mulx    a_0,b_2,t_1     !mul_add_c(a[0],b[2],c3,c1,c2);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(2)       !r[2]=c3;
+        or      c_12,c_3,c_12   !=
+
+        mulx    a_0,b_3,t_1     !mul_add_c(a[0],b[3],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        mulx    a_1,b_2,t_1     !mul_add_c(a[1],b[2],c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8        !=
+        add     c_3,t_2,c_3
+        lduw    ap(3),a_3
+        mulx    a_2,b_1,t_1     !mul_add_c(a[2],b[1],c1,c2,c3);
+        addcc   c_12,t_1,c_12   !=
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_3,b_0,t_1     !mul_add_c(a[3],b[0],c1,c2,c3);!=
+        addcc   c_12,t_1,t_1    !=
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(3)       !=!r[3]=c1;
+        or      c_12,c_3,c_12
+
+        mulx    a_3,b_1,t_1     !mul_add_c(a[3],b[1],c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3             !=
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_2,b_2,t_1     !mul_add_c(a[2],b[2],c2,c3,c1);
+        addcc   c_12,t_1,c_12   !=
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_1,b_3,t_1     !mul_add_c(a[1],b[3],c2,c3,c1);
+        addcc   c_12,t_1,t_1    !=
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(4)       !=!r[4]=c2;
+        or      c_12,c_3,c_12
+
+        mulx    a_2,b_3,t_1     !mul_add_c(a[2],b[3],c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        clr     c_3             !=
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_3,b_2,t_1     !mul_add_c(a[3],b[2],c3,c1,c2);
+        addcc   c_12,t_1,t_1    !=
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(5)       !=!r[5]=c3;
+        or      c_12,c_3,c_12
+
+        mulx    a_3,b_3,t_1     !mul_add_c(a[3],b[3],c1,c2,c3);
+        addcc   c_12,t_1,t_1
+        srlx    t_1,32,c_12     !=
+        stuw    t_1,rp(6)       !r[6]=c1;
+        stuw    c_12,rp(7)      !r[7]=c2;
+        
+        ret
+        restore %g0,%g0,%o0
+
+.type   bn_mul_comba4,#function
+.size   bn_mul_comba4,(.-bn_mul_comba4)
+
+.align  32
+
+.global bn_sqr_comba8
+bn_sqr_comba8:
+        save    %sp,FRAME_SIZE,%sp
+        mov     1,t_2
+        lduw    ap(0),a_0
+        sllx    t_2,32,t_2
+        lduw    ap(1),a_1
+        mulx    a_0,a_0,t_1     !sqr_add_c(a,0,c1,c2,c3);
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(0)       !r[0]=c1;
+
+        lduw    ap(2),a_2
+        mulx    a_0,a_1,t_1     !=!sqr_add_c2(a,1,0,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(1)       !r[1]=c2;
+        or      c_12,c_3,c_12
+
+        mulx    a_2,a_0,t_1     !sqr_add_c2(a,2,0,c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    ap(3),a_3
+        mulx    a_1,a_1,t_1     !sqr_add_c(a,1,c3,c1,c2);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(2)       !r[2]=c3;
+        or      c_12,c_3,c_12
+
+        mulx    a_0,a_3,t_1     !sqr_add_c2(a,3,0,c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    ap(4),a_4
+        mulx    a_1,a_2,t_1     !sqr_add_c2(a,2,1,c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        st      t_1,rp(3)       !r[3]=c1;
+        or      c_12,c_3,c_12
+
+        mulx    a_4,a_0,t_1     !sqr_add_c2(a,4,0,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_3,a_1,t_1     !sqr_add_c2(a,3,1,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    ap(5),a_5
+        mulx    a_2,a_2,t_1     !sqr_add_c(a,2,c2,c3,c1);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(4)       !r[4]=c2;
+        or      c_12,c_3,c_12
+
+        mulx    a_0,a_5,t_1     !sqr_add_c2(a,5,0,c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_1,a_4,t_1     !sqr_add_c2(a,4,1,c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    ap(6),a_6
+        mulx    a_2,a_3,t_1     !sqr_add_c2(a,3,2,c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(5)       !r[5]=c3;
+        or      c_12,c_3,c_12
+
+        mulx    a_6,a_0,t_1     !sqr_add_c2(a,6,0,c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_5,a_1,t_1     !sqr_add_c2(a,5,1,c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_4,a_2,t_1     !sqr_add_c2(a,4,2,c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    ap(7),a_7
+        mulx    a_3,a_3,t_1     !=!sqr_add_c(a,3,c1,c2,c3);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(6)       !r[6]=c1;
+        or      c_12,c_3,c_12
+
+        mulx    a_0,a_7,t_1     !sqr_add_c2(a,7,0,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_1,a_6,t_1     !sqr_add_c2(a,6,1,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_2,a_5,t_1     !sqr_add_c2(a,5,2,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_3,a_4,t_1     !sqr_add_c2(a,4,3,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(7)       !r[7]=c2;
+        or      c_12,c_3,c_12
+
+        mulx    a_7,a_1,t_1     !sqr_add_c2(a,7,1,c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_6,a_2,t_1     !sqr_add_c2(a,6,2,c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_5,a_3,t_1     !sqr_add_c2(a,5,3,c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_4,a_4,t_1     !sqr_add_c(a,4,c3,c1,c2);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(8)       !r[8]=c3;
+        or      c_12,c_3,c_12
+
+        mulx    a_2,a_7,t_1     !sqr_add_c2(a,7,2,c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_3,a_6,t_1     !sqr_add_c2(a,6,3,c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_4,a_5,t_1     !sqr_add_c2(a,5,4,c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(9)       !r[9]=c1;
+        or      c_12,c_3,c_12
+
+        mulx    a_7,a_3,t_1     !sqr_add_c2(a,7,3,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_6,a_4,t_1     !sqr_add_c2(a,6,4,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_5,a_5,t_1     !sqr_add_c(a,5,c2,c3,c1);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(10)      !r[10]=c2;
+        or      c_12,c_3,c_12
+
+        mulx    a_4,a_7,t_1     !sqr_add_c2(a,7,4,c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_5,a_6,t_1     !sqr_add_c2(a,6,5,c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(11)      !r[11]=c3;
+        or      c_12,c_3,c_12
+
+        mulx    a_7,a_5,t_1     !sqr_add_c2(a,7,5,c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_6,a_6,t_1     !sqr_add_c(a,6,c1,c2,c3);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(12)      !r[12]=c1;
+        or      c_12,c_3,c_12
+
+        mulx    a_6,a_7,t_1     !sqr_add_c2(a,7,6,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(13)      !r[13]=c2;
+        or      c_12,c_3,c_12
+
+        mulx    a_7,a_7,t_1     !sqr_add_c(a,7,c3,c1,c2);
+        addcc   c_12,t_1,t_1
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(14)      !r[14]=c3;
+        stuw    c_12,rp(15)     !r[15]=c1;
+
+        ret
+        restore %g0,%g0,%o0
+
+.type   bn_sqr_comba8,#function
+.size   bn_sqr_comba8,(.-bn_sqr_comba8)
+
+.align  32
+
+.global bn_sqr_comba4
+/*
+ * void bn_sqr_comba4(r,a)
+ * BN_ULONG *r,*a;
+ */
+bn_sqr_comba4:
+        save    %sp,FRAME_SIZE,%sp
+        mov     1,t_2
+        lduw    ap(0),a_0
+        sllx    t_2,32,t_2
+        lduw    ap(1),a_1
+        mulx    a_0,a_0,t_1     !sqr_add_c(a,0,c1,c2,c3);
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(0)       !r[0]=c1;
+
+        lduw    ap(2),a_2
+        mulx    a_0,a_1,t_1     !sqr_add_c2(a,1,0,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(1)       !r[1]=c2;
+        or      c_12,c_3,c_12
+
+        mulx    a_2,a_0,t_1     !sqr_add_c2(a,2,0,c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        lduw    ap(3),a_3
+        mulx    a_1,a_1,t_1     !sqr_add_c(a,1,c3,c1,c2);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(2)       !r[2]=c3;
+        or      c_12,c_3,c_12
+
+        mulx    a_0,a_3,t_1     !sqr_add_c2(a,3,0,c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_1,a_2,t_1     !sqr_add_c2(a,2,1,c1,c2,c3);
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(3)       !r[3]=c1;
+        or      c_12,c_3,c_12
+
+        mulx    a_3,a_1,t_1     !sqr_add_c2(a,3,1,c2,c3,c1);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,c_12
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        mulx    a_2,a_2,t_1     !sqr_add_c(a,2,c2,c3,c1);
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(4)       !r[4]=c2;
+        or      c_12,c_3,c_12
+
+        mulx    a_2,a_3,t_1     !sqr_add_c2(a,3,2,c3,c1,c2);
+        addcc   c_12,t_1,c_12
+        clr     c_3
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        addcc   c_12,t_1,t_1
+        bcs,a   %xcc,.+8
+        add     c_3,t_2,c_3
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(5)       !r[5]=c3;
+        or      c_12,c_3,c_12
+
+        mulx    a_3,a_3,t_1     !sqr_add_c(a,3,c1,c2,c3);
+        addcc   c_12,t_1,t_1
+        srlx    t_1,32,c_12
+        stuw    t_1,rp(6)       !r[6]=c1;
+        stuw    c_12,rp(7)      !r[7]=c2;
+        
+        ret
+        restore %g0,%g0,%o0
+
+.type   bn_sqr_comba4,#function
+.size   bn_sqr_comba4,(.-bn_sqr_comba4)
+
+.align  32
diff --git a/src/lib/libcrypto/bn/asm/vms.mar b/src/lib/libcrypto/bn/asm/vms.mar
new file mode 100644
index 0000000000..ac9d57d7b0
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/vms.mar
@@ -0,0 +1,6695 @@
+        .title  vax_bn_mul_add_word  unsigned multiply & add, 32*32+32+32=>64
+;
+; w.j.m. 15-jan-1999
+;
+; it's magic ...
+;
+; ULONG bn_mul_add_words(ULONG r[],ULONG a[],int n,ULONG w) {
+;       ULONG c = 0;
+;       int i;
+;       for(i = 0; i < n; i++) <c,r[i]> := r[i] + c + a[i] * w ;
+;       return c;
+; }
+
+r=4 ;(AP)
+a=8 ;(AP)
+n=12 ;(AP)      n       by value (input)
+w=16 ;(AP)      w       by value (input)
+
+
+        .psect  code,nowrt
+
+.entry  bn_mul_add_words,^m<r2,r3,r4,r5,r6>
+
+        moval   @r(ap),r2
+        moval   @a(ap),r3
+        movl    n(ap),r4        ; assumed >0 by C code
+        movl    w(ap),r5
+        clrl    r6              ; c
+
+0$:
+        emul    r5,(r3),(r2),r0         ; w, a[], r[] considered signed
+
+        ; fixup for "negative" r[]
+        tstl    (r2)
+        bgeq    10$
+        incl    r1
+10$:
+
+        ; add in c
+        addl2   r6,r0
+        adwc    #0,r1
+
+        ; combined fixup for "negative" w, a[]
+        tstl    r5
+        bgeq    20$
+        addl2   (r3),r1
+20$:
+        tstl    (r3)
+        bgeq    30$
+        addl2   r5,r1
+30$:
+
+        movl    r0,(r2)+                ; store lo result in r[] & advance
+        addl    #4,r3                   ; advance a[]
+        movl    r1,r6                   ; store hi result => c
+
+        sobgtr  r4,0$
+
+        movl    r6,r0                   ; return c
+        ret
+
+        .title  vax_bn_mul_word  unsigned multiply & add, 32*32+32=>64
+;
+; w.j.m. 15-jan-1999
+;
+; it's magic ...
+;
+; ULONG bn_mul_words(ULONG r[],ULONG a[],int n,ULONG w) {
+;       ULONG c = 0;
+;       int i;
+;       for(i = 0; i < num; i++) <c,r[i]> := a[i] * w + c ;
+;       return(c);
+; }
+
+r=4 ;(AP)
+a=8 ;(AP)
+n=12 ;(AP)      n       by value (input)
+w=16 ;(AP)      w       by value (input)
+
+
+        .psect  code,nowrt
+
+.entry  bn_mul_words,^m<r2,r3,r4,r5,r6>
+
+        moval   @r(ap),r2       ; r2 -> r[]
+        moval   @a(ap),r3       ; r3 -> a[]
+        movl    n(ap),r4        ; r4 = loop count (assumed >0 by C code)
+        movl    w(ap),r5        ; r5 = w
+        clrl    r6              ; r6 = c
+
+0$:
+        ; <r1,r0> := w * a[] + c
+        emul    r5,(r3),r6,r0           ; w, a[], c considered signed
+
+        ; fixup for "negative" c
+        tstl    r6                      ; c
+        bgeq    10$
+        incl    r1
+10$:
+
+        ; combined fixup for "negative" w, a[]
+        tstl    r5                      ; w
+        bgeq    20$
+        addl2   (r3),r1                 ; a[]
+20$:
+        tstl    (r3)                    ; a[]
+        bgeq    30$
+        addl2   r5,r1                   ; w
+30$:
+
+        movl    r0,(r2)+                ; store lo result in r[] & advance
+        addl    #4,r3                   ; advance a[]
+        movl    r1,r6                   ; store hi result => c
+
+        sobgtr  r4,0$
+
+        movl    r6,r0                   ; return c
+        ret
+
+        .title  vax_bn_sqr_words  unsigned square, 32*32=>64
+;
+; w.j.m. 15-jan-1999
+;
+; it's magic ...
+;
+; void bn_sqr_words(ULONG r[],ULONG a[],int n) {
+;       int i;
+;       for(i = 0; i < n; i++) <r[2*i+1],r[2*i]> := a[i] * a[i] ;
+; }
+
+r=4 ;(AP)
+a=8 ;(AP)
+n=12 ;(AP)      n       by value (input)
+
+
+        .psect  code,nowrt
+
+.entry  bn_sqr_words,^m<r2,r3,r4,r5>
+
+        moval   @r(ap),r2       ; r2 -> r[]
+        moval   @a(ap),r3       ; r3 -> a[]
+        movl    n(ap),r4        ; r4 = n (assumed >0 by C code)
+
+0$:
+        movl    (r3)+,r5                ; r5 = a[] & advance
+
+        ; <r1,r0> := a[] * a[]
+        emul    r5,r5,#0,r0             ; a[] considered signed
+
+        ; fixup for "negative" a[]
+        tstl    r5                      ; a[]
+        bgeq    30$
+        addl2   r5,r1                   ; a[]
+        addl2   r5,r1                   ; a[]
+30$:
+
+        movl    r0,(r2)+                ; store lo result in r[] & advance
+        movl    r1,(r2)+                ; store hi result in r[] & advance
+
+        sobgtr  r4,0$
+
+        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
+
+        ashl    #16,4(ap),r2
+        movzwl  2(r3),r0
+        bisl2   r0,r2
+        bicl3   #0,r2,4(ap)
+
+        bicl3   #-65536,(r3),r0
+        ashl    #16,r0,(r3)
+        brw     noname.7
+        nop     
+noname.18:
+
+        bisl2   r4,r9
+
+        movl    r9,r10
+
+noname.3:
+        movl    r10,r0
+        ret     
+        tstl    r0
+
+
+        .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
+
+        movl    r1,12(r3)
+
+        decl    16(ap)
+        bleq    noname.25
+noname.28:
+
+        addl3   #16,r6,8(ap)
+
+        addl3   #16,r5,12(ap)
+
+        addl3   #16,r3,4(ap)
+        brw     noname.23
+        tstl    r0
+noname.25:
+
+        movl    r4,r7
+
+noname.22:
+        movl    r7,r0
+        ret     
+        nop     
+
+
+
+;r=4 ;(AP)
+;a=8 ;(AP)
+;b=12 ;(AP)
+;n=16 ;(AP)     n       by value (input)
+
+        .psect  code,nowrt
+
+.entry  BN_SUB_WORDS,^m<r2,r3,r4,r5,r6,r7>
+
+        clrl    r6
+
+        tstl    16(ap)
+        bgtr    noname.31
+        clrl    r7
+        brw     noname.32
+        tstl    r0
+noname.31:
+
+noname.33:
+
+        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     
+
+
+
+;r=4 ;(AP)
+;a=8 ;(AP)
+;b=12 ;(AP)
+;n=16 ;(AP)     n       by value (input)
+
+        .psect  code,nowrt
+
+.entry  BN_MUL_COMBA8,^m<r2,r3,r4,r5,r6,r7,r8,r9,r10,r11>
+        movab   -924(sp),sp
+        clrq    r8
+
+        clrl    r10
+
+        movl    8(ap),r6
+        movzwl  2(r6),r3
+        movl    12(ap),r7
+        bicl3   #-65536,(r7),r2
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,(r6),-12(fp)
+        bicl3   #-65536,r3,-16(fp)
+        mull3   r0,-12(fp),-4(fp)
+        mull2   r2,-12(fp)
+        mull3   r2,-16(fp),-8(fp)
+        mull2   r0,-16(fp)
+        addl3   -4(fp),-8(fp),r0
+        bicl3   #0,r0,-4(fp)
+        cmpl    -4(fp),-8(fp)
+        bgequ   noname.45
+        addl2   #65536,-16(fp)
+noname.45:
+        movzwl  -2(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-16(fp)
+        bicl3   #-65536,-4(fp),r0
+        ashl    #16,r0,-8(fp)
+        addl3   -8(fp),-12(fp),r0
+        bicl3   #0,r0,-12(fp)
+        cmpl    -12(fp),-8(fp)
+        bgequ   noname.46
+        incl    -16(fp)
+noname.46:
+        movl    -12(fp),r1
+        movl    -16(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.47
+        incl    r2
+noname.47:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.48
+        incl    r10
+noname.48:
+
+        movl    4(ap),r11
+        movl    r9,(r11)
+
+        clrl    r9
+
+        movzwl  2(r6),r2
+        bicl3   #-65536,4(r7),r3
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,(r6),-28(fp)
+        bicl3   #-65536,r2,-32(fp)
+        mull3   r0,-28(fp),-20(fp)
+        mull2   r3,-28(fp)
+        mull3   r3,-32(fp),-24(fp)
+        mull2   r0,-32(fp)
+        addl3   -20(fp),-24(fp),r0
+        bicl3   #0,r0,-20(fp)
+        cmpl    -20(fp),-24(fp)
+        bgequ   noname.49
+        addl2   #65536,-32(fp)
+noname.49:
+        movzwl  -18(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-32(fp)
+        bicl3   #-65536,-20(fp),r0
+        ashl    #16,r0,-24(fp)
+        addl3   -24(fp),-28(fp),r0
+        bicl3   #0,r0,-28(fp)
+        cmpl    -28(fp),-24(fp)
+        bgequ   noname.50
+        incl    -32(fp)
+noname.50:
+        movl    -28(fp),r1
+        movl    -32(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.51
+        incl    r2
+noname.51:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.52
+        incl    r9
+noname.52:
+
+        movzwl  6(r6),r2
+        bicl3   #-65536,(r7),r3
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,4(r6),-44(fp)
+        bicl3   #-65536,r2,-48(fp)
+        mull3   r0,-44(fp),-36(fp)
+        mull2   r3,-44(fp)
+        mull3   r3,-48(fp),-40(fp)
+        mull2   r0,-48(fp)
+        addl3   -36(fp),-40(fp),r0
+        bicl3   #0,r0,-36(fp)
+        cmpl    -36(fp),-40(fp)
+        bgequ   noname.53
+        addl2   #65536,-48(fp)
+noname.53:
+        movzwl  -34(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-48(fp)
+        bicl3   #-65536,-36(fp),r0
+        ashl    #16,r0,-40(fp)
+        addl3   -40(fp),-44(fp),r0
+        bicl3   #0,r0,-44(fp)
+        cmpl    -44(fp),-40(fp)
+        bgequ   noname.54
+        incl    -48(fp)
+noname.54:
+        movl    -44(fp),r1
+        movl    -48(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.55
+        incl    r2
+noname.55:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.56
+        incl    r9
+noname.56:
+
+        movl    r8,4(r11)
+
+        clrl    r8
+
+        movzwl  10(r6),r2
+        bicl3   #-65536,(r7),r3
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,8(r6),-60(fp)
+        bicl3   #-65536,r2,-64(fp)
+        mull3   r0,-60(fp),-52(fp)
+        mull2   r3,-60(fp)
+        mull3   r3,-64(fp),-56(fp)
+        mull2   r0,-64(fp)
+        addl3   -52(fp),-56(fp),r0
+        bicl3   #0,r0,-52(fp)
+        cmpl    -52(fp),-56(fp)
+        bgequ   noname.57
+        addl2   #65536,-64(fp)
+noname.57:
+        movzwl  -50(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-64(fp)
+        bicl3   #-65536,-52(fp),r0
+        ashl    #16,r0,-56(fp)
+        addl3   -56(fp),-60(fp),r0
+        bicl3   #0,r0,-60(fp)
+        cmpl    -60(fp),-56(fp)
+        bgequ   noname.58
+        incl    -64(fp)
+noname.58:
+        movl    -60(fp),r1
+        movl    -64(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.59
+        incl    r2
+noname.59:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.60
+        incl    r8
+noname.60:
+
+        movzwl  6(r6),r2
+        bicl3   #-65536,4(r7),r3
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,4(r6),-76(fp)
+        bicl3   #-65536,r2,-80(fp)
+        mull3   r0,-76(fp),-68(fp)
+        mull2   r3,-76(fp)
+        mull3   r3,-80(fp),-72(fp)
+        mull2   r0,-80(fp)
+        addl3   -68(fp),-72(fp),r0
+        bicl3   #0,r0,-68(fp)
+        cmpl    -68(fp),-72(fp)
+        bgequ   noname.61
+        addl2   #65536,-80(fp)
+noname.61:
+        movzwl  -66(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-80(fp)
+        bicl3   #-65536,-68(fp),r0
+        ashl    #16,r0,-72(fp)
+        addl3   -72(fp),-76(fp),r0
+        bicl3   #0,r0,-76(fp)
+        cmpl    -76(fp),-72(fp)
+        bgequ   noname.62
+        incl    -80(fp)
+noname.62:
+        movl    -76(fp),r1
+        movl    -80(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.63
+        incl    r2
+noname.63:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.64
+        incl    r8
+noname.64:
+
+        movzwl  2(r6),r2
+        bicl3   #-65536,8(r7),r3
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,(r6),-92(fp)
+        bicl3   #-65536,r2,-96(fp)
+        mull3   r0,-92(fp),-84(fp)
+        mull2   r3,-92(fp)
+        mull3   r3,-96(fp),-88(fp)
+        mull2   r0,-96(fp)
+        addl3   -84(fp),-88(fp),r0
+        bicl3   #0,r0,-84(fp)
+        cmpl    -84(fp),-88(fp)
+        bgequ   noname.65
+        addl2   #65536,-96(fp)
+noname.65:
+        movzwl  -82(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-96(fp)
+        bicl3   #-65536,-84(fp),r0
+        ashl    #16,r0,-88(fp)
+        addl3   -88(fp),-92(fp),r0
+        bicl3   #0,r0,-92(fp)
+        cmpl    -92(fp),-88(fp)
+        bgequ   noname.66
+        incl    -96(fp)
+noname.66:
+        movl    -92(fp),r1
+        movl    -96(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.67
+        incl    r2
+noname.67:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.68
+        incl    r8
+noname.68:
+
+        movl    r10,8(r11)
+
+        clrl    r10
+
+        movzwl  2(r6),r2
+        bicl3   #-65536,12(r7),r3
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,(r6),-108(fp)
+        bicl3   #-65536,r2,-112(fp)
+        mull3   r0,-108(fp),-100(fp)
+        mull2   r3,-108(fp)
+        mull3   r3,-112(fp),-104(fp)
+        mull2   r0,-112(fp)
+        addl3   -100(fp),-104(fp),r0
+        bicl3   #0,r0,-100(fp)
+        cmpl    -100(fp),-104(fp)
+        bgequ   noname.69
+        addl2   #65536,-112(fp)
+noname.69:
+        movzwl  -98(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-112(fp)
+        bicl3   #-65536,-100(fp),r0
+        ashl    #16,r0,-104(fp)
+        addl3   -104(fp),-108(fp),r0
+        bicl3   #0,r0,-108(fp)
+        cmpl    -108(fp),-104(fp)
+        bgequ   noname.70
+        incl    -112(fp)
+noname.70:
+        movl    -108(fp),r1
+        movl    -112(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.71
+        incl    r2
+noname.71:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.72
+        incl    r10
+noname.72:
+
+        movzwl  6(r6),r2
+        bicl3   #-65536,8(r7),r3
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,4(r6),-124(fp)
+        bicl3   #-65536,r2,-128(fp)
+        mull3   r0,-124(fp),-116(fp)
+        mull2   r3,-124(fp)
+        mull3   r3,-128(fp),-120(fp)
+        mull2   r0,-128(fp)
+        addl3   -116(fp),-120(fp),r0
+        bicl3   #0,r0,-116(fp)
+        cmpl    -116(fp),-120(fp)
+        bgequ   noname.73
+        addl2   #65536,-128(fp)
+noname.73:
+        movzwl  -114(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-128(fp)
+        bicl3   #-65536,-116(fp),r0
+        ashl    #16,r0,-120(fp)
+        addl3   -120(fp),-124(fp),r0
+        bicl3   #0,r0,-124(fp)
+        cmpl    -124(fp),-120(fp)
+        bgequ   noname.74
+        incl    -128(fp)
+noname.74:
+        movl    -124(fp),r1
+        movl    -128(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.75
+        incl    r2
+noname.75:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.76
+        incl    r10
+noname.76:
+
+        movzwl  10(r6),r2
+        bicl3   #-65536,4(r7),r3
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,8(r6),-140(fp)
+        bicl3   #-65536,r2,-144(fp)
+        mull3   r0,-140(fp),-132(fp)
+        mull2   r3,-140(fp)
+        mull3   r3,-144(fp),-136(fp)
+        mull2   r0,-144(fp)
+        addl3   -132(fp),-136(fp),r0
+        bicl3   #0,r0,-132(fp)
+        cmpl    -132(fp),-136(fp)
+        bgequ   noname.77
+        addl2   #65536,-144(fp)
+noname.77:
+        movzwl  -130(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-144(fp)
+        bicl3   #-65536,-132(fp),r0
+        ashl    #16,r0,-136(fp)
+        addl3   -136(fp),-140(fp),r0
+        bicl3   #0,r0,-140(fp)
+        cmpl    -140(fp),-136(fp)
+        bgequ   noname.78
+        incl    -144(fp)
+noname.78:
+        movl    -140(fp),r1
+        movl    -144(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.79
+        incl    r2
+noname.79:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.80
+        incl    r10
+noname.80:
+
+        movzwl  14(r6),r2
+        bicl3   #-65536,(r7),r3
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,12(r6),-156(fp)
+        bicl3   #-65536,r2,-160(fp)
+        mull3   r0,-156(fp),-148(fp)
+        mull2   r3,-156(fp)
+        mull3   r3,-160(fp),-152(fp)
+        mull2   r0,-160(fp)
+        addl3   -148(fp),-152(fp),r0
+        bicl3   #0,r0,-148(fp)
+        cmpl    -148(fp),-152(fp)
+        bgequ   noname.81
+        addl2   #65536,-160(fp)
+noname.81:
+        movzwl  -146(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-160(fp)
+        bicl3   #-65536,-148(fp),r0
+        ashl    #16,r0,-152(fp)
+        addl3   -152(fp),-156(fp),r0
+        bicl3   #0,r0,-156(fp)
+        cmpl    -156(fp),-152(fp)
+        bgequ   noname.82
+        incl    -160(fp)
+noname.82:
+        movl    -156(fp),r1
+        movl    -160(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.83
+        incl    r2
+noname.83:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.84
+        incl    r10
+noname.84:
+
+        movl    r9,12(r11)
+
+        clrl    r9
+
+        movzwl  18(r6),r2
+        bicl3   #-65536,(r7),r3
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,16(r6),-172(fp)
+        bicl3   #-65536,r2,-176(fp)
+        mull3   r0,-172(fp),-164(fp)
+        mull2   r3,-172(fp)
+        mull3   r3,-176(fp),-168(fp)
+        mull2   r0,-176(fp)
+        addl3   -164(fp),-168(fp),r0
+        bicl3   #0,r0,-164(fp)
+        cmpl    -164(fp),-168(fp)
+        bgequ   noname.85
+        addl2   #65536,-176(fp)
+noname.85:
+        movzwl  -162(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-176(fp)
+        bicl3   #-65536,-164(fp),r0
+        ashl    #16,r0,-168(fp)
+        addl3   -168(fp),-172(fp),r0
+        bicl3   #0,r0,-172(fp)
+        cmpl    -172(fp),-168(fp)
+        bgequ   noname.86
+        incl    -176(fp)
+noname.86:
+        movl    -172(fp),r1
+        movl    -176(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.87
+        incl    r2
+noname.87:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.88
+        incl    r9
+noname.88:
+
+        movzwl  14(r6),r2
+        bicl3   #-65536,4(r7),r3
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,12(r6),-188(fp)
+        bicl3   #-65536,r2,-192(fp)
+        mull3   r0,-188(fp),-180(fp)
+        mull2   r3,-188(fp)
+        mull3   r3,-192(fp),-184(fp)
+        mull2   r0,-192(fp)
+        addl3   -180(fp),-184(fp),r0
+        bicl3   #0,r0,-180(fp)
+        cmpl    -180(fp),-184(fp)
+        bgequ   noname.89
+        addl2   #65536,-192(fp)
+noname.89:
+        movzwl  -178(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-192(fp)
+        bicl3   #-65536,-180(fp),r0
+        ashl    #16,r0,-184(fp)
+        addl3   -184(fp),-188(fp),r0
+        bicl3   #0,r0,-188(fp)
+        cmpl    -188(fp),-184(fp)
+        bgequ   noname.90
+        incl    -192(fp)
+noname.90:
+        movl    -188(fp),r1
+        movl    -192(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.91
+        incl    r2
+noname.91:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.92
+        incl    r9
+noname.92:
+
+        movzwl  10(r6),r2
+        bicl3   #-65536,8(r7),r3
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,8(r6),-204(fp)
+        bicl3   #-65536,r2,-208(fp)
+        mull3   r0,-204(fp),-196(fp)
+        mull2   r3,-204(fp)
+        mull3   r3,-208(fp),-200(fp)
+        mull2   r0,-208(fp)
+        addl3   -196(fp),-200(fp),r0
+        bicl3   #0,r0,-196(fp)
+        cmpl    -196(fp),-200(fp)
+        bgequ   noname.93
+        addl2   #65536,-208(fp)
+noname.93:
+        movzwl  -194(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-208(fp)
+        bicl3   #-65536,-196(fp),r0
+        ashl    #16,r0,-200(fp)
+        addl3   -200(fp),-204(fp),r0
+        bicl3   #0,r0,-204(fp)
+        cmpl    -204(fp),-200(fp)
+        bgequ   noname.94
+        incl    -208(fp)
+noname.94:
+        movl    -204(fp),r1
+        movl    -208(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.95
+        incl    r2
+noname.95:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.96
+        incl    r9
+noname.96:
+
+        movzwl  6(r6),r2
+        bicl3   #-65536,12(r7),r3
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,4(r6),-220(fp)
+        bicl3   #-65536,r2,-224(fp)
+        mull3   r0,-220(fp),-212(fp)
+        mull2   r3,-220(fp)
+        mull3   r3,-224(fp),-216(fp)
+        mull2   r0,-224(fp)
+        addl3   -212(fp),-216(fp),r0
+        bicl3   #0,r0,-212(fp)
+        cmpl    -212(fp),-216(fp)
+        bgequ   noname.97
+        addl2   #65536,-224(fp)
+noname.97:
+        movzwl  -210(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-224(fp)
+        bicl3   #-65536,-212(fp),r0
+        ashl    #16,r0,-216(fp)
+        addl3   -216(fp),-220(fp),r0
+        bicl3   #0,r0,-220(fp)
+        cmpl    -220(fp),-216(fp)
+        bgequ   noname.98
+        incl    -224(fp)
+noname.98:
+        movl    -220(fp),r1
+        movl    -224(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.99
+        incl    r2
+noname.99:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.100
+        incl    r9
+noname.100:
+
+        movzwl  2(r6),r2
+        bicl3   #-65536,16(r7),r3
+        movzwl  18(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,(r6),-236(fp)
+        bicl3   #-65536,r2,-240(fp)
+        mull3   r0,-236(fp),-228(fp)
+        mull2   r3,-236(fp)
+        mull3   r3,-240(fp),-232(fp)
+        mull2   r0,-240(fp)
+        addl3   -228(fp),-232(fp),r0
+        bicl3   #0,r0,-228(fp)
+        cmpl    -228(fp),-232(fp)
+        bgequ   noname.101
+        addl2   #65536,-240(fp)
+noname.101:
+        movzwl  -226(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-240(fp)
+        bicl3   #-65536,-228(fp),r0
+        ashl    #16,r0,-232(fp)
+        addl3   -232(fp),-236(fp),r0
+        bicl3   #0,r0,-236(fp)
+        cmpl    -236(fp),-232(fp)
+        bgequ   noname.102
+        incl    -240(fp)
+noname.102:
+        movl    -236(fp),r1
+        movl    -240(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.103
+        incl    r2
+noname.103:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.104
+        incl    r9
+noname.104:
+
+        movl    r8,16(r11)
+
+        clrl    r8
+
+        movzwl  2(r6),r2
+        bicl3   #-65536,20(r7),r3
+        movzwl  22(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,(r6),-252(fp)
+        bicl3   #-65536,r2,-256(fp)
+        mull3   r0,-252(fp),-244(fp)
+        mull2   r3,-252(fp)
+        mull3   r3,-256(fp),-248(fp)
+        mull2   r0,-256(fp)
+        addl3   -244(fp),-248(fp),r0
+        bicl3   #0,r0,-244(fp)
+        cmpl    -244(fp),-248(fp)
+        bgequ   noname.105
+        addl2   #65536,-256(fp)
+noname.105:
+        movzwl  -242(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-256(fp)
+        bicl3   #-65536,-244(fp),r0
+        ashl    #16,r0,-248(fp)
+        addl3   -248(fp),-252(fp),r0
+        bicl3   #0,r0,-252(fp)
+        cmpl    -252(fp),-248(fp)
+        bgequ   noname.106
+        incl    -256(fp)
+noname.106:
+        movl    -252(fp),r1
+        movl    -256(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.107
+        incl    r2
+noname.107:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.108
+        incl    r8
+noname.108:
+
+        movzwl  6(r6),r2
+        bicl3   #-65536,16(r7),r3
+        movzwl  18(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,4(r6),-268(fp)
+        bicl3   #-65536,r2,-272(fp)
+        mull3   r0,-268(fp),-260(fp)
+        mull2   r3,-268(fp)
+        mull3   r3,-272(fp),-264(fp)
+        mull2   r0,-272(fp)
+        addl3   -260(fp),-264(fp),r0
+        bicl3   #0,r0,-260(fp)
+        cmpl    -260(fp),-264(fp)
+        bgequ   noname.109
+        addl2   #65536,-272(fp)
+noname.109:
+        movzwl  -258(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-272(fp)
+        bicl3   #-65536,-260(fp),r0
+        ashl    #16,r0,-264(fp)
+        addl3   -264(fp),-268(fp),r0
+        bicl3   #0,r0,-268(fp)
+        cmpl    -268(fp),-264(fp)
+        bgequ   noname.110
+        incl    -272(fp)
+noname.110:
+        movl    -268(fp),r1
+        movl    -272(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.111
+        incl    r2
+noname.111:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.112
+        incl    r8
+noname.112:
+
+        movzwl  10(r6),r2
+        bicl3   #-65536,12(r7),r3
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,8(r6),-284(fp)
+        bicl3   #-65536,r2,-288(fp)
+        mull3   r0,-284(fp),-276(fp)
+        mull2   r3,-284(fp)
+        mull3   r3,-288(fp),-280(fp)
+        mull2   r0,-288(fp)
+        addl3   -276(fp),-280(fp),r0
+        bicl3   #0,r0,-276(fp)
+        cmpl    -276(fp),-280(fp)
+        bgequ   noname.113
+        addl2   #65536,-288(fp)
+noname.113:
+        movzwl  -274(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-288(fp)
+        bicl3   #-65536,-276(fp),r0
+        ashl    #16,r0,-280(fp)
+        addl3   -280(fp),-284(fp),r0
+        bicl3   #0,r0,-284(fp)
+        cmpl    -284(fp),-280(fp)
+        bgequ   noname.114
+        incl    -288(fp)
+noname.114:
+        movl    -284(fp),r1
+        movl    -288(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.115
+        incl    r2
+noname.115:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.116
+        incl    r8
+noname.116:
+
+        movzwl  14(r6),r2
+        bicl3   #-65536,8(r7),r3
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,12(r6),-300(fp)
+        bicl3   #-65536,r2,-304(fp)
+        mull3   r0,-300(fp),-292(fp)
+        mull2   r3,-300(fp)
+        mull3   r3,-304(fp),-296(fp)
+        mull2   r0,-304(fp)
+        addl3   -292(fp),-296(fp),r0
+        bicl3   #0,r0,-292(fp)
+        cmpl    -292(fp),-296(fp)
+        bgequ   noname.117
+        addl2   #65536,-304(fp)
+noname.117:
+        movzwl  -290(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-304(fp)
+        bicl3   #-65536,-292(fp),r0
+        ashl    #16,r0,-296(fp)
+        addl3   -296(fp),-300(fp),r0
+        bicl3   #0,r0,-300(fp)
+        cmpl    -300(fp),-296(fp)
+        bgequ   noname.118
+        incl    -304(fp)
+noname.118:
+        movl    -300(fp),r1
+        movl    -304(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.119
+        incl    r2
+noname.119:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.120
+        incl    r8
+noname.120:
+
+        movzwl  18(r6),r2
+        bicl3   #-65536,4(r7),r3
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,16(r6),-316(fp)
+        bicl3   #-65536,r2,-320(fp)
+        mull3   r0,-316(fp),-308(fp)
+        mull2   r3,-316(fp)
+        mull3   r3,-320(fp),-312(fp)
+        mull2   r0,-320(fp)
+        addl3   -308(fp),-312(fp),r0
+        bicl3   #0,r0,-308(fp)
+        cmpl    -308(fp),-312(fp)
+        bgequ   noname.121
+        addl2   #65536,-320(fp)
+noname.121:
+        movzwl  -306(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-320(fp)
+        bicl3   #-65536,-308(fp),r0
+        ashl    #16,r0,-312(fp)
+        addl3   -312(fp),-316(fp),r0
+        bicl3   #0,r0,-316(fp)
+        cmpl    -316(fp),-312(fp)
+        bgequ   noname.122
+        incl    -320(fp)
+noname.122:
+        movl    -316(fp),r1
+        movl    -320(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.123
+        incl    r2
+
+noname.123:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.124
+        incl    r8
+noname.124:
+
+        movzwl  22(r6),r2
+        bicl3   #-65536,(r7),r3
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,20(r6),-332(fp)
+        bicl3   #-65536,r2,-336(fp)
+        mull3   r0,-332(fp),-324(fp)
+        mull2   r3,-332(fp)
+        mull3   r3,-336(fp),-328(fp)
+        mull2   r0,-336(fp)
+        addl3   -324(fp),-328(fp),r0
+        bicl3   #0,r0,-324(fp)
+        cmpl    -324(fp),-328(fp)
+        bgequ   noname.125
+        addl2   #65536,-336(fp)
+noname.125:
+        movzwl  -322(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-336(fp)
+        bicl3   #-65536,-324(fp),r0
+        ashl    #16,r0,-328(fp)
+        addl3   -328(fp),-332(fp),r0
+        bicl3   #0,r0,-332(fp)
+        cmpl    -332(fp),-328(fp)
+        bgequ   noname.126
+        incl    -336(fp)
+noname.126:
+        movl    -332(fp),r1
+        movl    -336(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.127
+        incl    r2
+noname.127:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.128
+        incl    r8
+noname.128:
+
+        movl    r10,20(r11)
+
+        clrl    r10
+
+        movzwl  26(r6),r2
+        bicl3   #-65536,(r7),r3
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,24(r6),-348(fp)
+        bicl3   #-65536,r2,-352(fp)
+        mull3   r0,-348(fp),-340(fp)
+        mull2   r3,-348(fp)
+        mull3   r3,-352(fp),-344(fp)
+        mull2   r0,-352(fp)
+        addl3   -340(fp),-344(fp),r0
+        bicl3   #0,r0,-340(fp)
+        cmpl    -340(fp),-344(fp)
+        bgequ   noname.129
+        addl2   #65536,-352(fp)
+noname.129:
+        movzwl  -338(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-352(fp)
+        bicl3   #-65536,-340(fp),r0
+        ashl    #16,r0,-344(fp)
+        addl3   -344(fp),-348(fp),r0
+        bicl3   #0,r0,-348(fp)
+        cmpl    -348(fp),-344(fp)
+        bgequ   noname.130
+        incl    -352(fp)
+noname.130:
+        movl    -348(fp),r1
+        movl    -352(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.131
+        incl    r2
+noname.131:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.132
+        incl    r10
+noname.132:
+
+        movzwl  22(r6),r2
+        bicl3   #-65536,4(r7),r3
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,20(r6),-364(fp)
+        bicl3   #-65536,r2,-368(fp)
+        mull3   r0,-364(fp),-356(fp)
+        mull2   r3,-364(fp)
+        mull3   r3,-368(fp),-360(fp)
+        mull2   r0,-368(fp)
+        addl3   -356(fp),-360(fp),r0
+        bicl3   #0,r0,-356(fp)
+        cmpl    -356(fp),-360(fp)
+        bgequ   noname.133
+        addl2   #65536,-368(fp)
+noname.133:
+        movzwl  -354(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-368(fp)
+        bicl3   #-65536,-356(fp),r0
+        ashl    #16,r0,-360(fp)
+        addl3   -360(fp),-364(fp),r0
+        bicl3   #0,r0,-364(fp)
+        cmpl    -364(fp),-360(fp)
+        bgequ   noname.134
+        incl    -368(fp)
+noname.134:
+        movl    -364(fp),r1
+        movl    -368(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.135
+        incl    r2
+noname.135:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.136
+        incl    r10
+noname.136:
+
+        movzwl  18(r6),r2
+        bicl3   #-65536,8(r7),r3
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,16(r6),-380(fp)
+        bicl3   #-65536,r2,-384(fp)
+        mull3   r0,-380(fp),-372(fp)
+        mull2   r3,-380(fp)
+        mull3   r3,-384(fp),-376(fp)
+        mull2   r0,-384(fp)
+        addl3   -372(fp),-376(fp),r0
+        bicl3   #0,r0,-372(fp)
+        cmpl    -372(fp),-376(fp)
+        bgequ   noname.137
+        addl2   #65536,-384(fp)
+noname.137:
+        movzwl  -370(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-384(fp)
+        bicl3   #-65536,-372(fp),r0
+        ashl    #16,r0,-376(fp)
+        addl3   -376(fp),-380(fp),r0
+        bicl3   #0,r0,-380(fp)
+        cmpl    -380(fp),-376(fp)
+        bgequ   noname.138
+        incl    -384(fp)
+noname.138:
+        movl    -380(fp),r1
+        movl    -384(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.139
+        incl    r2
+noname.139:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.140
+        incl    r10
+noname.140:
+
+        movzwl  14(r6),r2
+        bicl3   #-65536,12(r7),r3
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,12(r6),-396(fp)
+        bicl3   #-65536,r2,-400(fp)
+        mull3   r0,-396(fp),-388(fp)
+        mull2   r3,-396(fp)
+        mull3   r3,-400(fp),-392(fp)
+        mull2   r0,-400(fp)
+        addl3   -388(fp),-392(fp),r0
+        bicl3   #0,r0,-388(fp)
+        cmpl    -388(fp),-392(fp)
+        bgequ   noname.141
+        addl2   #65536,-400(fp)
+noname.141:
+        movzwl  -386(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-400(fp)
+        bicl3   #-65536,-388(fp),r0
+        ashl    #16,r0,-392(fp)
+        addl3   -392(fp),-396(fp),r0
+        bicl3   #0,r0,-396(fp)
+        cmpl    -396(fp),-392(fp)
+        bgequ   noname.142
+        incl    -400(fp)
+noname.142:
+        movl    -396(fp),r1
+        movl    -400(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.143
+        incl    r2
+noname.143:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.144
+        incl    r10
+noname.144:
+
+        movzwl  10(r6),r2
+        bicl3   #-65536,16(r7),r3
+        movzwl  18(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,8(r6),-412(fp)
+        bicl3   #-65536,r2,-416(fp)
+        mull3   r0,-412(fp),-404(fp)
+        mull2   r3,-412(fp)
+        mull3   r3,-416(fp),-408(fp)
+        mull2   r0,-416(fp)
+        addl3   -404(fp),-408(fp),r0
+        bicl3   #0,r0,-404(fp)
+        cmpl    -404(fp),-408(fp)
+        bgequ   noname.145
+        addl2   #65536,-416(fp)
+noname.145:
+        movzwl  -402(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-416(fp)
+        bicl3   #-65536,-404(fp),r0
+        ashl    #16,r0,-408(fp)
+        addl3   -408(fp),-412(fp),r0
+        bicl3   #0,r0,-412(fp)
+        cmpl    -412(fp),-408(fp)
+        bgequ   noname.146
+        incl    -416(fp)
+noname.146:
+        movl    -412(fp),r1
+        movl    -416(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.147
+        incl    r2
+noname.147:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.148
+        incl    r10
+noname.148:
+
+        movzwl  6(r6),r2
+        bicl3   #-65536,20(r7),r3
+        movzwl  22(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,4(r6),-428(fp)
+        bicl3   #-65536,r2,-432(fp)
+        mull3   r0,-428(fp),-420(fp)
+        mull2   r3,-428(fp)
+        mull3   r3,-432(fp),-424(fp)
+        mull2   r0,-432(fp)
+        addl3   -420(fp),-424(fp),r0
+        bicl3   #0,r0,-420(fp)
+        cmpl    -420(fp),-424(fp)
+        bgequ   noname.149
+        addl2   #65536,-432(fp)
+noname.149:
+        movzwl  -418(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-432(fp)
+        bicl3   #-65536,-420(fp),r0
+        ashl    #16,r0,-424(fp)
+        addl3   -424(fp),-428(fp),r0
+        bicl3   #0,r0,-428(fp)
+        cmpl    -428(fp),-424(fp)
+        bgequ   noname.150
+        incl    -432(fp)
+noname.150:
+        movl    -428(fp),r1
+        movl    -432(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.151
+        incl    r2
+noname.151:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.152
+        incl    r10
+noname.152:
+
+        movzwl  2(r6),r2
+        bicl3   #-65536,24(r7),r3
+        movzwl  26(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,(r6),-444(fp)
+        bicl3   #-65536,r2,-448(fp)
+        mull3   r0,-444(fp),-436(fp)
+        mull2   r3,-444(fp)
+        mull3   r3,-448(fp),-440(fp)
+        mull2   r0,-448(fp)
+        addl3   -436(fp),-440(fp),r0
+        bicl3   #0,r0,-436(fp)
+        cmpl    -436(fp),-440(fp)
+        bgequ   noname.153
+        addl2   #65536,-448(fp)
+noname.153:
+        movzwl  -434(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-448(fp)
+        bicl3   #-65536,-436(fp),r0
+        ashl    #16,r0,-440(fp)
+        addl3   -440(fp),-444(fp),r0
+        bicl3   #0,r0,-444(fp)
+        cmpl    -444(fp),-440(fp)
+        bgequ   noname.154
+        incl    -448(fp)
+noname.154:
+        movl    -444(fp),r1
+        movl    -448(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.155
+        incl    r2
+noname.155:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.156
+        incl    r10
+noname.156:
+
+        movl    r9,24(r11)
+
+        clrl    r9
+
+        movzwl  2(r6),r2
+        bicl3   #-65536,28(r7),r3
+        movzwl  30(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,(r6),-460(fp)
+        bicl3   #-65536,r2,-464(fp)
+        mull3   r0,-460(fp),-452(fp)
+        mull2   r3,-460(fp)
+        mull3   r3,-464(fp),-456(fp)
+        mull2   r0,-464(fp)
+        addl3   -452(fp),-456(fp),r0
+        bicl3   #0,r0,-452(fp)
+        cmpl    -452(fp),-456(fp)
+        bgequ   noname.157
+        addl2   #65536,-464(fp)
+noname.157:
+        movzwl  -450(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-464(fp)
+        bicl3   #-65536,-452(fp),r0
+        ashl    #16,r0,-456(fp)
+        addl3   -456(fp),-460(fp),r0
+        bicl3   #0,r0,-460(fp)
+        cmpl    -460(fp),-456(fp)
+        bgequ   noname.158
+        incl    -464(fp)
+noname.158:
+        movl    -460(fp),r1
+        movl    -464(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.159
+        incl    r2
+noname.159:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.160
+        incl    r9
+noname.160:
+
+        movzwl  6(r6),r2
+        bicl3   #-65536,24(r7),r3
+        movzwl  26(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,4(r6),-476(fp)
+        bicl3   #-65536,r2,-480(fp)
+        mull3   r0,-476(fp),-468(fp)
+        mull2   r3,-476(fp)
+        mull3   r3,-480(fp),-472(fp)
+        mull2   r0,-480(fp)
+        addl3   -468(fp),-472(fp),r0
+        bicl3   #0,r0,-468(fp)
+        cmpl    -468(fp),-472(fp)
+        bgequ   noname.161
+        addl2   #65536,-480(fp)
+noname.161:
+        movzwl  -466(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-480(fp)
+        bicl3   #-65536,-468(fp),r0
+        ashl    #16,r0,-472(fp)
+        addl3   -472(fp),-476(fp),r0
+        bicl3   #0,r0,-476(fp)
+        cmpl    -476(fp),-472(fp)
+        bgequ   noname.162
+        incl    -480(fp)
+noname.162:
+        movl    -476(fp),r1
+        movl    -480(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.163
+        incl    r2
+noname.163:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.164
+        incl    r9
+noname.164:
+
+        movzwl  10(r6),r2
+        bicl3   #-65536,20(r7),r3
+        movzwl  22(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,8(r6),-492(fp)
+        bicl3   #-65536,r2,-496(fp)
+        mull3   r0,-492(fp),-484(fp)
+        mull2   r3,-492(fp)
+        mull3   r3,-496(fp),-488(fp)
+        mull2   r0,-496(fp)
+        addl3   -484(fp),-488(fp),r0
+        bicl3   #0,r0,-484(fp)
+        cmpl    -484(fp),-488(fp)
+        bgequ   noname.165
+        addl2   #65536,-496(fp)
+noname.165:
+        movzwl  -482(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-496(fp)
+        bicl3   #-65536,-484(fp),r0
+        ashl    #16,r0,-488(fp)
+        addl3   -488(fp),-492(fp),r0
+        bicl3   #0,r0,-492(fp)
+        cmpl    -492(fp),-488(fp)
+        bgequ   noname.166
+        incl    -496(fp)
+noname.166:
+        movl    -492(fp),r1
+        movl    -496(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.167
+        incl    r2
+noname.167:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.168
+        incl    r9
+noname.168:
+
+        movzwl  14(r6),r2
+        bicl3   #-65536,16(r7),r3
+        movzwl  18(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,12(r6),-508(fp)
+        bicl3   #-65536,r2,-512(fp)
+        mull3   r0,-508(fp),-500(fp)
+        mull2   r3,-508(fp)
+        mull3   r3,-512(fp),-504(fp)
+        mull2   r0,-512(fp)
+        addl3   -500(fp),-504(fp),r0
+        bicl3   #0,r0,-500(fp)
+        cmpl    -500(fp),-504(fp)
+        bgequ   noname.169
+        addl2   #65536,-512(fp)
+noname.169:
+        movzwl  -498(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-512(fp)
+        bicl3   #-65536,-500(fp),r0
+        ashl    #16,r0,-504(fp)
+        addl3   -504(fp),-508(fp),r0
+        bicl3   #0,r0,-508(fp)
+        cmpl    -508(fp),-504(fp)
+        bgequ   noname.170
+        incl    -512(fp)
+noname.170:
+        movl    -508(fp),r1
+        movl    -512(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.171
+        incl    r2
+noname.171:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.172
+        incl    r9
+noname.172:
+
+        movzwl  18(r6),r2
+        bicl3   #-65536,12(r7),r3
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,16(r6),-524(fp)
+        bicl3   #-65536,r2,-528(fp)
+        mull3   r0,-524(fp),-516(fp)
+        mull2   r3,-524(fp)
+        mull3   r3,-528(fp),-520(fp)
+        mull2   r0,-528(fp)
+        addl3   -516(fp),-520(fp),r0
+        bicl3   #0,r0,-516(fp)
+        cmpl    -516(fp),-520(fp)
+        bgequ   noname.173
+        addl2   #65536,-528(fp)
+noname.173:
+        movzwl  -514(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-528(fp)
+        bicl3   #-65536,-516(fp),r0
+        ashl    #16,r0,-520(fp)
+        addl3   -520(fp),-524(fp),r0
+        bicl3   #0,r0,-524(fp)
+        cmpl    -524(fp),-520(fp)
+        bgequ   noname.174
+        incl    -528(fp)
+noname.174:
+        movl    -524(fp),r1
+        movl    -528(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.175
+        incl    r2
+noname.175:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.176
+        incl    r9
+noname.176:
+
+        movzwl  22(r6),r2
+        bicl3   #-65536,8(r7),r3
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,20(r6),-540(fp)
+        bicl3   #-65536,r2,-544(fp)
+        mull3   r0,-540(fp),-532(fp)
+        mull2   r3,-540(fp)
+        mull3   r3,-544(fp),-536(fp)
+        mull2   r0,-544(fp)
+        addl3   -532(fp),-536(fp),r0
+        bicl3   #0,r0,-532(fp)
+        cmpl    -532(fp),-536(fp)
+        bgequ   noname.177
+        addl2   #65536,-544(fp)
+noname.177:
+        movzwl  -530(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-544(fp)
+        bicl3   #-65536,-532(fp),r0
+        ashl    #16,r0,-536(fp)
+        addl3   -536(fp),-540(fp),r0
+        bicl3   #0,r0,-540(fp)
+        cmpl    -540(fp),-536(fp)
+        bgequ   noname.178
+        incl    -544(fp)
+noname.178:
+        movl    -540(fp),r1
+        movl    -544(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.179
+        incl    r2
+noname.179:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.180
+        incl    r9
+noname.180:
+
+        movzwl  26(r6),r2
+        bicl3   #-65536,4(r7),r3
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,24(r6),-556(fp)
+        bicl3   #-65536,r2,-560(fp)
+        mull3   r0,-556(fp),-548(fp)
+        mull2   r3,-556(fp)
+        mull3   r3,-560(fp),-552(fp)
+        mull2   r0,-560(fp)
+        addl3   -548(fp),-552(fp),r0
+        bicl3   #0,r0,-548(fp)
+        cmpl    -548(fp),-552(fp)
+        bgequ   noname.181
+        addl2   #65536,-560(fp)
+noname.181:
+        movzwl  -546(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-560(fp)
+        bicl3   #-65536,-548(fp),r0
+        ashl    #16,r0,-552(fp)
+        addl3   -552(fp),-556(fp),r0
+        bicl3   #0,r0,-556(fp)
+        cmpl    -556(fp),-552(fp)
+        bgequ   noname.182
+        incl    -560(fp)
+noname.182:
+        movl    -556(fp),r1
+        movl    -560(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.183
+        incl    r2
+noname.183:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.184
+        incl    r9
+noname.184:
+
+        movzwl  30(r6),r2
+        bicl3   #-65536,(r7),r3
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,28(r6),-572(fp)
+        bicl3   #-65536,r2,-576(fp)
+        mull3   r0,-572(fp),-564(fp)
+        mull2   r3,-572(fp)
+        mull3   r3,-576(fp),-568(fp)
+        mull2   r0,-576(fp)
+        addl3   -564(fp),-568(fp),r0
+        bicl3   #0,r0,-564(fp)
+        cmpl    -564(fp),-568(fp)
+        bgequ   noname.185
+        addl2   #65536,-576(fp)
+noname.185:
+        movzwl  -562(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-576(fp)
+        bicl3   #-65536,-564(fp),r0
+        ashl    #16,r0,-568(fp)
+        addl3   -568(fp),-572(fp),r0
+        bicl3   #0,r0,-572(fp)
+        cmpl    -572(fp),-568(fp)
+        bgequ   noname.186
+        incl    -576(fp)
+noname.186:
+        movl    -572(fp),r1
+        movl    -576(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.187
+        incl    r2
+noname.187:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.188
+        incl    r9
+noname.188:
+
+        movl    r8,28(r11)
+
+        clrl    r8
+
+        movzwl  30(r6),r2
+        bicl3   #-65536,4(r7),r3
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,28(r6),-588(fp)
+        bicl3   #-65536,r2,-592(fp)
+        mull3   r0,-588(fp),-580(fp)
+        mull2   r3,-588(fp)
+        mull3   r3,-592(fp),-584(fp)
+        mull2   r0,-592(fp)
+        addl3   -580(fp),-584(fp),r0
+        bicl3   #0,r0,-580(fp)
+        cmpl    -580(fp),-584(fp)
+        bgequ   noname.189
+        addl2   #65536,-592(fp)
+noname.189:
+        movzwl  -578(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-592(fp)
+        bicl3   #-65536,-580(fp),r0
+        ashl    #16,r0,-584(fp)
+        addl3   -584(fp),-588(fp),r0
+        bicl3   #0,r0,-588(fp)
+        cmpl    -588(fp),-584(fp)
+        bgequ   noname.190
+        incl    -592(fp)
+noname.190:
+        movl    -588(fp),r1
+        movl    -592(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.191
+        incl    r2
+noname.191:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.192
+        incl    r8
+noname.192:
+
+        movzwl  26(r6),r2
+        bicl3   #-65536,8(r7),r3
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,24(r6),-604(fp)
+        bicl3   #-65536,r2,-608(fp)
+        mull3   r0,-604(fp),-596(fp)
+        mull2   r3,-604(fp)
+        mull3   r3,-608(fp),-600(fp)
+        mull2   r0,-608(fp)
+        addl3   -596(fp),-600(fp),r0
+        bicl3   #0,r0,-596(fp)
+        cmpl    -596(fp),-600(fp)
+        bgequ   noname.193
+        addl2   #65536,-608(fp)
+noname.193:
+        movzwl  -594(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-608(fp)
+        bicl3   #-65536,-596(fp),r0
+        ashl    #16,r0,-600(fp)
+        addl3   -600(fp),-604(fp),r0
+        bicl3   #0,r0,-604(fp)
+        cmpl    -604(fp),-600(fp)
+        bgequ   noname.194
+        incl    -608(fp)
+noname.194:
+        movl    -604(fp),r1
+        movl    -608(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.195
+        incl    r2
+noname.195:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.196
+        incl    r8
+noname.196:
+
+        movzwl  22(r6),r2
+        bicl3   #-65536,12(r7),r3
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,20(r6),-620(fp)
+        bicl3   #-65536,r2,-624(fp)
+        mull3   r0,-620(fp),-612(fp)
+        mull2   r3,-620(fp)
+        mull3   r3,-624(fp),-616(fp)
+        mull2   r0,-624(fp)
+        addl3   -612(fp),-616(fp),r0
+        bicl3   #0,r0,-612(fp)
+        cmpl    -612(fp),-616(fp)
+        bgequ   noname.197
+        addl2   #65536,-624(fp)
+noname.197:
+        movzwl  -610(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-624(fp)
+        bicl3   #-65536,-612(fp),r0
+        ashl    #16,r0,-616(fp)
+        addl3   -616(fp),-620(fp),r0
+        bicl3   #0,r0,-620(fp)
+        cmpl    -620(fp),-616(fp)
+        bgequ   noname.198
+        incl    -624(fp)
+noname.198:
+        movl    -620(fp),r1
+        movl    -624(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.199
+        incl    r2
+noname.199:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.200
+        incl    r8
+noname.200:
+
+        movzwl  18(r6),r2
+        bicl3   #-65536,16(r7),r3
+        movzwl  18(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,16(r6),-636(fp)
+        bicl3   #-65536,r2,-640(fp)
+        mull3   r0,-636(fp),-628(fp)
+        mull2   r3,-636(fp)
+        mull3   r3,-640(fp),-632(fp)
+        mull2   r0,-640(fp)
+        addl3   -628(fp),-632(fp),r0
+        bicl3   #0,r0,-628(fp)
+        cmpl    -628(fp),-632(fp)
+        bgequ   noname.201
+        addl2   #65536,-640(fp)
+noname.201:
+        movzwl  -626(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-640(fp)
+        bicl3   #-65536,-628(fp),r0
+        ashl    #16,r0,-632(fp)
+        addl3   -632(fp),-636(fp),r0
+        bicl3   #0,r0,-636(fp)
+        cmpl    -636(fp),-632(fp)
+        bgequ   noname.202
+        incl    -640(fp)
+noname.202:
+        movl    -636(fp),r1
+        movl    -640(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.203
+        incl    r2
+noname.203:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.204
+        incl    r8
+noname.204:
+
+        movzwl  14(r6),r2
+        bicl3   #-65536,20(r7),r3
+        movzwl  22(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,12(r6),-652(fp)
+        bicl3   #-65536,r2,-656(fp)
+        mull3   r0,-652(fp),-644(fp)
+        mull2   r3,-652(fp)
+        mull3   r3,-656(fp),-648(fp)
+        mull2   r0,-656(fp)
+        addl3   -644(fp),-648(fp),r0
+        bicl3   #0,r0,-644(fp)
+        cmpl    -644(fp),-648(fp)
+        bgequ   noname.205
+        addl2   #65536,-656(fp)
+noname.205:
+        movzwl  -642(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-656(fp)
+        bicl3   #-65536,-644(fp),r0
+        ashl    #16,r0,-648(fp)
+        addl3   -648(fp),-652(fp),r0
+        bicl3   #0,r0,-652(fp)
+        cmpl    -652(fp),-648(fp)
+        bgequ   noname.206
+        incl    -656(fp)
+noname.206:
+        movl    -652(fp),r1
+        movl    -656(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.207
+        incl    r2
+noname.207:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.208
+        incl    r8
+noname.208:
+
+        movzwl  10(r6),r2
+        bicl3   #-65536,24(r7),r3
+        movzwl  26(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,8(r6),-668(fp)
+        bicl3   #-65536,r2,-672(fp)
+        mull3   r0,-668(fp),-660(fp)
+        mull2   r3,-668(fp)
+        mull3   r3,-672(fp),-664(fp)
+        mull2   r0,-672(fp)
+        addl3   -660(fp),-664(fp),r0
+        bicl3   #0,r0,-660(fp)
+        cmpl    -660(fp),-664(fp)
+        bgequ   noname.209
+        addl2   #65536,-672(fp)
+noname.209:
+        movzwl  -658(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-672(fp)
+        bicl3   #-65536,-660(fp),r0
+        ashl    #16,r0,-664(fp)
+        addl3   -664(fp),-668(fp),r0
+        bicl3   #0,r0,-668(fp)
+        cmpl    -668(fp),-664(fp)
+        bgequ   noname.210
+        incl    -672(fp)
+noname.210:
+        movl    -668(fp),r1
+        movl    -672(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.211
+        incl    r2
+noname.211:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.212
+        incl    r8
+noname.212:
+
+        movzwl  6(r6),r2
+        bicl3   #-65536,28(r7),r3
+        movzwl  30(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,4(r6),-684(fp)
+        bicl3   #-65536,r2,-688(fp)
+        mull3   r0,-684(fp),-676(fp)
+        mull2   r3,-684(fp)
+        mull3   r3,-688(fp),-680(fp)
+        mull2   r0,-688(fp)
+        addl3   -676(fp),-680(fp),r0
+        bicl3   #0,r0,-676(fp)
+        cmpl    -676(fp),-680(fp)
+        bgequ   noname.213
+        addl2   #65536,-688(fp)
+noname.213:
+        movzwl  -674(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-688(fp)
+        bicl3   #-65536,-676(fp),r0
+        ashl    #16,r0,-680(fp)
+        addl3   -680(fp),-684(fp),r0
+        bicl3   #0,r0,-684(fp)
+        cmpl    -684(fp),-680(fp)
+        bgequ   noname.214
+        incl    -688(fp)
+noname.214:
+        movl    -684(fp),r1
+        movl    -688(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.215
+        incl    r2
+noname.215:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.216
+        incl    r8
+noname.216:
+
+        movl    r10,32(r11)
+
+        clrl    r10
+
+        movzwl  10(r6),r2
+        bicl3   #-65536,28(r7),r3
+        movzwl  30(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,8(r6),-700(fp)
+        bicl3   #-65536,r2,-704(fp)
+        mull3   r0,-700(fp),-692(fp)
+        mull2   r3,-700(fp)
+        mull3   r3,-704(fp),-696(fp)
+        mull2   r0,-704(fp)
+        addl3   -692(fp),-696(fp),r0
+        bicl3   #0,r0,-692(fp)
+        cmpl    -692(fp),-696(fp)
+        bgequ   noname.217
+        addl2   #65536,-704(fp)
+noname.217:
+        movzwl  -690(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-704(fp)
+        bicl3   #-65536,-692(fp),r0
+        ashl    #16,r0,-696(fp)
+        addl3   -696(fp),-700(fp),r0
+        bicl3   #0,r0,-700(fp)
+        cmpl    -700(fp),-696(fp)
+        bgequ   noname.218
+        incl    -704(fp)
+noname.218:
+        movl    -700(fp),r1
+        movl    -704(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.219
+        incl    r2
+noname.219:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.220
+        incl    r10
+noname.220:
+
+        movzwl  14(r6),r2
+        bicl3   #-65536,24(r7),r3
+        movzwl  26(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,12(r6),-716(fp)
+        bicl3   #-65536,r2,-720(fp)
+        mull3   r0,-716(fp),-708(fp)
+        mull2   r3,-716(fp)
+        mull3   r3,-720(fp),-712(fp)
+        mull2   r0,-720(fp)
+        addl3   -708(fp),-712(fp),r0
+        bicl3   #0,r0,-708(fp)
+        cmpl    -708(fp),-712(fp)
+        bgequ   noname.221
+        addl2   #65536,-720(fp)
+noname.221:
+        movzwl  -706(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-720(fp)
+        bicl3   #-65536,-708(fp),r0
+        ashl    #16,r0,-712(fp)
+        addl3   -712(fp),-716(fp),r0
+        bicl3   #0,r0,-716(fp)
+        cmpl    -716(fp),-712(fp)
+        bgequ   noname.222
+        incl    -720(fp)
+noname.222:
+        movl    -716(fp),r1
+        movl    -720(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.223
+        incl    r2
+noname.223:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.224
+        incl    r10
+noname.224:
+
+        movzwl  18(r6),r2
+        bicl3   #-65536,20(r7),r3
+        movzwl  22(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,16(r6),-732(fp)
+        bicl3   #-65536,r2,-736(fp)
+        mull3   r0,-732(fp),-724(fp)
+        mull2   r3,-732(fp)
+        mull3   r3,-736(fp),-728(fp)
+        mull2   r0,-736(fp)
+        addl3   -724(fp),-728(fp),r0
+        bicl3   #0,r0,-724(fp)
+        cmpl    -724(fp),-728(fp)
+        bgequ   noname.225
+        addl2   #65536,-736(fp)
+noname.225:
+        movzwl  -722(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-736(fp)
+        bicl3   #-65536,-724(fp),r0
+        ashl    #16,r0,-728(fp)
+        addl3   -728(fp),-732(fp),r0
+        bicl3   #0,r0,-732(fp)
+        cmpl    -732(fp),-728(fp)
+        bgequ   noname.226
+        incl    -736(fp)
+noname.226:
+        movl    -732(fp),r1
+        movl    -736(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.227
+        incl    r2
+noname.227:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.228
+        incl    r10
+noname.228:
+
+        movzwl  22(r6),r2
+        bicl3   #-65536,16(r7),r3
+        movzwl  18(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,20(r6),-748(fp)
+        bicl3   #-65536,r2,-752(fp)
+        mull3   r0,-748(fp),-740(fp)
+        mull2   r3,-748(fp)
+        mull3   r3,-752(fp),-744(fp)
+        mull2   r0,-752(fp)
+        addl3   -740(fp),-744(fp),r0
+        bicl3   #0,r0,-740(fp)
+        cmpl    -740(fp),-744(fp)
+        bgequ   noname.229
+        addl2   #65536,-752(fp)
+noname.229:
+        movzwl  -738(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-752(fp)
+        bicl3   #-65536,-740(fp),r0
+        ashl    #16,r0,-744(fp)
+        addl3   -744(fp),-748(fp),r0
+        bicl3   #0,r0,-748(fp)
+        cmpl    -748(fp),-744(fp)
+        bgequ   noname.230
+        incl    -752(fp)
+noname.230:
+        movl    -748(fp),r1
+        movl    -752(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.231
+        incl    r2
+noname.231:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.232
+        incl    r10
+noname.232:
+
+        movzwl  26(r6),r2
+        bicl3   #-65536,12(r7),r3
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,24(r6),-764(fp)
+        bicl3   #-65536,r2,-768(fp)
+        mull3   r0,-764(fp),-756(fp)
+        mull2   r3,-764(fp)
+        mull3   r3,-768(fp),-760(fp)
+        mull2   r0,-768(fp)
+        addl3   -756(fp),-760(fp),r0
+        bicl3   #0,r0,-756(fp)
+        cmpl    -756(fp),-760(fp)
+        bgequ   noname.233
+        addl2   #65536,-768(fp)
+noname.233:
+        movzwl  -754(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-768(fp)
+        bicl3   #-65536,-756(fp),r0
+        ashl    #16,r0,-760(fp)
+        addl3   -760(fp),-764(fp),r0
+        bicl3   #0,r0,-764(fp)
+        cmpl    -764(fp),-760(fp)
+        bgequ   noname.234
+        incl    -768(fp)
+noname.234:
+        movl    -764(fp),r1
+        movl    -768(fp),r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.235
+        incl    r2
+noname.235:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.236
+        incl    r10
+noname.236:
+
+        bicl3   #-65536,28(r6),r3
+        movzwl  30(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,8(r7),r2
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-772(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-776(fp)
+        mull2   r0,r4
+        addl3   -772(fp),-776(fp),r0
+        bicl3   #0,r0,-772(fp)
+        cmpl    -772(fp),-776(fp)
+        bgequ   noname.237
+        addl2   #65536,r4
+noname.237:
+        movzwl  -770(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-772(fp),r0
+        ashl    #16,r0,-776(fp)
+        addl2   -776(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-776(fp)
+        bgequ   noname.238
+        incl    r4
+noname.238:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.239
+        incl    r2
+noname.239:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.240
+        incl    r10
+noname.240:
+
+        movl    r9,36(r11)
+
+        clrl    r9
+
+        bicl3   #-65536,28(r6),r3
+        movzwl  30(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,12(r7),r2
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-780(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-784(fp)
+        mull2   r0,r4
+        addl3   -780(fp),-784(fp),r0
+        bicl3   #0,r0,-780(fp)
+        cmpl    -780(fp),-784(fp)
+        bgequ   noname.241
+        addl2   #65536,r4
+noname.241:
+        movzwl  -778(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-780(fp),r0
+        ashl    #16,r0,-784(fp)
+        addl2   -784(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-784(fp)
+        bgequ   noname.242
+        incl    r4
+noname.242:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.243
+        incl    r2
+noname.243:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.244
+        incl    r9
+noname.244:
+
+        bicl3   #-65536,24(r6),r3
+        movzwl  26(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,16(r7),r2
+        movzwl  18(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-788(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-792(fp)
+        mull2   r0,r4
+        addl3   -788(fp),-792(fp),r0
+        bicl3   #0,r0,-788(fp)
+        cmpl    -788(fp),-792(fp)
+        bgequ   noname.245
+        addl2   #65536,r4
+noname.245:
+        movzwl  -786(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-788(fp),r0
+        ashl    #16,r0,-792(fp)
+        addl2   -792(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-792(fp)
+        bgequ   noname.246
+        incl    r4
+noname.246:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.247
+        incl    r2
+noname.247:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.248
+        incl    r9
+noname.248:
+
+        bicl3   #-65536,20(r6),r3
+        movzwl  22(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,20(r7),r2
+        movzwl  22(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-796(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-800(fp)
+        mull2   r0,r4
+        addl3   -796(fp),-800(fp),r0
+        bicl3   #0,r0,-796(fp)
+        cmpl    -796(fp),-800(fp)
+        bgequ   noname.249
+        addl2   #65536,r4
+noname.249:
+        movzwl  -794(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-796(fp),r0
+        ashl    #16,r0,-800(fp)
+        addl2   -800(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-800(fp)
+        bgequ   noname.250
+        incl    r4
+noname.250:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.251
+        incl    r2
+noname.251:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.252
+        incl    r9
+noname.252:
+
+        bicl3   #-65536,16(r6),r3
+        movzwl  18(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,24(r7),r2
+        movzwl  26(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-804(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-808(fp)
+        mull2   r0,r4
+        addl3   -804(fp),-808(fp),r0
+        bicl3   #0,r0,-804(fp)
+        cmpl    -804(fp),-808(fp)
+        bgequ   noname.253
+        addl2   #65536,r4
+noname.253:
+        movzwl  -802(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-804(fp),r0
+        ashl    #16,r0,-808(fp)
+        addl2   -808(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-808(fp)
+        bgequ   noname.254
+        incl    r4
+noname.254:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.255
+        incl    r2
+noname.255:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.256
+        incl    r9
+noname.256:
+
+        bicl3   #-65536,12(r6),r3
+        movzwl  14(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,28(r7),r2
+        movzwl  30(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-812(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-816(fp)
+        mull2   r0,r4
+        addl3   -812(fp),-816(fp),r0
+        bicl3   #0,r0,-812(fp)
+        cmpl    -812(fp),-816(fp)
+        bgequ   noname.257
+        addl2   #65536,r4
+noname.257:
+        movzwl  -810(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-812(fp),r0
+        ashl    #16,r0,-816(fp)
+        addl2   -816(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-816(fp)
+        bgequ   noname.258
+        incl    r4
+noname.258:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.259
+        incl    r2
+noname.259:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.260
+        incl    r9
+noname.260:
+
+        movl    r8,40(r11)
+
+        clrl    r8
+
+        bicl3   #-65536,16(r6),r3
+        movzwl  18(r6),r2
+        bicl3   #-65536,28(r7),r1
+        movzwl  30(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r4
+        bicl3   #-65536,r2,-828(fp)
+        mull3   r0,r4,-820(fp)
+        mull2   r1,r4
+        mull3   r1,-828(fp),-824(fp)
+        mull2   r0,-828(fp)
+        addl3   -820(fp),-824(fp),r0
+        bicl3   #0,r0,-820(fp)
+        cmpl    -820(fp),-824(fp)
+        bgequ   noname.261
+        addl2   #65536,-828(fp)
+noname.261:
+        movzwl  -818(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-828(fp)
+        bicl3   #-65536,-820(fp),r0
+        ashl    #16,r0,-824(fp)
+        addl2   -824(fp),r4
+        bicl2   #0,r4
+        cmpl    r4,-824(fp)
+        bgequ   noname.262
+        incl    -828(fp)
+noname.262:
+        movl    r4,r1
+        movl    -828(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.263
+        incl    r2
+noname.263:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.264
+        incl    r8
+noname.264:
+
+        movzwl  22(r6),r2
+        bicl3   #-65536,24(r7),r3
+        movzwl  26(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,20(r6),-840(fp)
+        bicl3   #-65536,r2,-844(fp)
+        mull3   r0,-840(fp),-832(fp)
+        mull2   r3,-840(fp)
+        mull3   r3,-844(fp),-836(fp)
+        mull2   r0,-844(fp)
+        addl3   -832(fp),-836(fp),r0
+        bicl3   #0,r0,-832(fp)
+        cmpl    -832(fp),-836(fp)
+        bgequ   noname.265
+        addl2   #65536,-844(fp)
+noname.265:
+        movzwl  -830(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-844(fp)
+        bicl3   #-65536,-832(fp),r0
+        ashl    #16,r0,-836(fp)
+        addl3   -836(fp),-840(fp),r0
+        bicl3   #0,r0,-840(fp)
+        cmpl    -840(fp),-836(fp)
+        bgequ   noname.266
+        incl    -844(fp)
+noname.266:
+        movl    -840(fp),r1
+        movl    -844(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.267
+        incl    r2
+noname.267:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.268
+        incl    r8
+noname.268:
+
+        bicl3   #-65536,24(r6),r3
+        movzwl  26(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,20(r7),r2
+        movzwl  22(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-848(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-852(fp)
+        mull2   r0,r4
+        addl3   -848(fp),-852(fp),r0
+        bicl3   #0,r0,-848(fp)
+        cmpl    -848(fp),-852(fp)
+        bgequ   noname.269
+        addl2   #65536,r4
+noname.269:
+        movzwl  -846(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-848(fp),r0
+        ashl    #16,r0,-852(fp)
+        addl2   -852(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-852(fp)
+        bgequ   noname.270
+        incl    r4
+noname.270:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.271
+        incl    r2
+noname.271:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.272
+        incl    r8
+noname.272:
+
+        bicl3   #-65536,28(r6),r3
+        movzwl  30(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,16(r7),r2
+        movzwl  18(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-856(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-860(fp)
+        mull2   r0,r4
+        addl3   -856(fp),-860(fp),r0
+        bicl3   #0,r0,-856(fp)
+        cmpl    -856(fp),-860(fp)
+        bgequ   noname.273
+        addl2   #65536,r4
+noname.273:
+        movzwl  -854(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-856(fp),r0
+        ashl    #16,r0,-860(fp)
+        addl2   -860(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-860(fp)
+        bgequ   noname.274
+        incl    r4
+noname.274:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.275
+        incl    r2
+noname.275:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.276
+        incl    r8
+noname.276:
+
+        movl    r10,44(r11)
+
+        clrl    r10
+
+        bicl3   #-65536,28(r6),r3
+        movzwl  30(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,20(r7),r2
+        movzwl  22(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-864(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-868(fp)
+        mull2   r0,r4
+        addl3   -864(fp),-868(fp),r0
+        bicl3   #0,r0,-864(fp)
+        cmpl    -864(fp),-868(fp)
+        bgequ   noname.277
+        addl2   #65536,r4
+noname.277:
+        movzwl  -862(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-864(fp),r0
+        ashl    #16,r0,-868(fp)
+        addl2   -868(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-868(fp)
+        bgequ   noname.278
+        incl    r4
+noname.278:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.279
+        incl    r2
+noname.279:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.280
+        incl    r10
+noname.280:
+
+        bicl3   #-65536,24(r6),r3
+        movzwl  26(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,24(r7),r2
+        movzwl  26(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-872(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-876(fp)
+        mull2   r0,r4
+        addl3   -872(fp),-876(fp),r0
+        bicl3   #0,r0,-872(fp)
+        cmpl    -872(fp),-876(fp)
+        bgequ   noname.281
+        addl2   #65536,r4
+noname.281:
+        movzwl  -870(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-872(fp),r0
+        ashl    #16,r0,-876(fp)
+        addl2   -876(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-876(fp)
+        bgequ   noname.282
+        incl    r4
+noname.282:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.283
+        incl    r2
+noname.283:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.284
+        incl    r10
+noname.284:
+
+        bicl3   #-65536,20(r6),r3
+        movzwl  22(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,28(r7),r2
+        movzwl  30(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-880(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-884(fp)
+        mull2   r0,r4
+        addl3   -880(fp),-884(fp),r0
+        bicl3   #0,r0,-880(fp)
+        cmpl    -880(fp),-884(fp)
+        bgequ   noname.285
+        addl2   #65536,r4
+noname.285:
+        movzwl  -878(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-880(fp),r0
+        ashl    #16,r0,-884(fp)
+        addl2   -884(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-884(fp)
+        bgequ   noname.286
+        incl    r4
+noname.286:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.287
+        incl    r2
+noname.287:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.288
+        incl    r10
+noname.288:
+
+        movl    r9,48(r11)
+
+        clrl    r9
+
+        bicl3   #-65536,24(r6),r3
+        movzwl  26(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,28(r7),r2
+        movzwl  30(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-888(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-892(fp)
+        mull2   r0,r4
+        addl3   -888(fp),-892(fp),r0
+        bicl3   #0,r0,-888(fp)
+        cmpl    -888(fp),-892(fp)
+        bgequ   noname.289
+        addl2   #65536,r4
+noname.289:
+        movzwl  -886(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-888(fp),r0
+        ashl    #16,r0,-892(fp)
+        addl2   -892(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-892(fp)
+        bgequ   noname.290
+        incl    r4
+noname.290:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.291
+        incl    r2
+noname.291:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.292
+        incl    r9
+noname.292:
+
+        movzwl  30(r6),r2
+        bicl3   #-65536,24(r7),r3
+        movzwl  26(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,28(r6),-904(fp)
+        bicl3   #-65536,r2,-908(fp)
+        mull3   r0,-904(fp),-896(fp)
+        mull2   r3,-904(fp)
+        mull3   r3,-908(fp),-900(fp)
+        mull2   r0,-908(fp)
+        addl3   -896(fp),-900(fp),r0
+        bicl3   #0,r0,-896(fp)
+        cmpl    -896(fp),-900(fp)
+        bgequ   noname.293
+        addl2   #65536,-908(fp)
+noname.293:
+        movzwl  -894(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-908(fp)
+        bicl3   #-65536,-896(fp),r0
+        ashl    #16,r0,-900(fp)
+        addl3   -900(fp),-904(fp),r0
+        bicl3   #0,r0,-904(fp)
+        cmpl    -904(fp),-900(fp)
+        bgequ   noname.294
+        incl    -908(fp)
+noname.294:
+        movl    -904(fp),r1
+        movl    -908(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.295
+        incl    r2
+noname.295:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.296
+        incl    r9
+noname.296:
+
+        movl    r8,52(r11)
+
+        clrl    r8
+
+        movzwl  30(r6),r2
+        bicl3   #-65536,28(r7),r3
+        movzwl  30(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,28(r6),-920(fp)
+        bicl3   #-65536,r2,-924(fp)
+        mull3   r0,-920(fp),-912(fp)
+        mull2   r3,-920(fp)
+        mull3   r3,-924(fp),-916(fp)
+        mull2   r0,-924(fp)
+        addl3   -912(fp),-916(fp),r0
+        bicl3   #0,r0,-912(fp)
+        cmpl    -912(fp),-916(fp)
+        bgequ   noname.297
+        addl2   #65536,-924(fp)
+noname.297:
+        movzwl  -910(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-924(fp)
+        bicl3   #-65536,-912(fp),r0
+        ashl    #16,r0,-916(fp)
+        addl3   -916(fp),-920(fp),r0
+        bicl3   #0,r0,-920(fp)
+        cmpl    -920(fp),-916(fp)
+        bgequ   noname.298
+        incl    -924(fp)
+noname.298:
+        movl    -920(fp),r1
+        movl    -924(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.299
+        incl    r2
+noname.299:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.300
+        incl    r8
+noname.300:
+
+        movl    r10,56(r11)
+
+        movl    r9,60(r11)
+
+        ret     
+
+
+
+;r=4 ;(AP)
+;a=8 ;(AP)
+;b=12 ;(AP)
+;n=16 ;(AP)     n       by value (input)
+
+        .psect  code,nowrt
+
+.entry  BN_MUL_COMBA4,^m<r2,r3,r4,r5,r6,r7,r8,r9,r10,r11>
+        movab   -156(sp),sp
+
+        clrq    r9
+
+        clrl    r8
+
+        movl    8(ap),r6
+        bicl3   #-65536,(r6),r3
+        movzwl  2(r6),r2
+        bicl2   #-65536,r2
+        movl    12(ap),r7
+        bicl3   #-65536,(r7),r1
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r2,r4
+        mull3   r0,r5,-4(fp)
+        mull2   r1,r5
+        mull3   r1,r4,-8(fp)
+        mull2   r0,r4
+        addl3   -4(fp),-8(fp),r0
+        bicl3   #0,r0,-4(fp)
+        cmpl    -4(fp),-8(fp)
+        bgequ   noname.303
+        addl2   #65536,r4
+noname.303:
+        movzwl  -2(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-4(fp),r0
+        ashl    #16,r0,-8(fp)
+        addl2   -8(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-8(fp)
+        bgequ   noname.304
+        incl    r4
+noname.304:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.305
+        incl    r2
+noname.305:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.306
+        incl    r8
+noname.306:
+
+        movl    4(ap),r11
+        movl    r10,(r11)
+
+        clrl    r10
+
+        bicl3   #-65536,(r6),r3
+        movzwl  2(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,4(r7),r2
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-12(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-16(fp)
+        mull2   r0,r4
+        addl3   -12(fp),-16(fp),r0
+        bicl3   #0,r0,-12(fp)
+        cmpl    -12(fp),-16(fp)
+        bgequ   noname.307
+        addl2   #65536,r4
+noname.307:
+        movzwl  -10(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-12(fp),r0
+        ashl    #16,r0,-16(fp)
+        addl2   -16(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-16(fp)
+        bgequ   noname.308
+        incl    r4
+noname.308:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.309
+        incl    r2
+noname.309:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.310
+        incl    r10
+noname.310:
+
+        bicl3   #-65536,4(r6),r3
+        movzwl  6(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,(r7),r2
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-20(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-24(fp)
+        mull2   r0,r4
+        addl3   -20(fp),-24(fp),r0
+        bicl3   #0,r0,-20(fp)
+        cmpl    -20(fp),-24(fp)
+        bgequ   noname.311
+        addl2   #65536,r4
+noname.311:
+        movzwl  -18(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-20(fp),r0
+        ashl    #16,r0,-24(fp)
+        addl2   -24(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-24(fp)
+        bgequ   noname.312
+        incl    r4
+noname.312:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.313
+        incl    r2
+noname.313:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.314
+        incl    r10
+noname.314:
+
+        movl    r9,4(r11)
+
+        clrl    r9
+
+        bicl3   #-65536,8(r6),r3
+        movzwl  10(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,(r7),r2
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-28(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-32(fp)
+        mull2   r0,r4
+        addl3   -28(fp),-32(fp),r0
+        bicl3   #0,r0,-28(fp)
+        cmpl    -28(fp),-32(fp)
+        bgequ   noname.315
+        addl2   #65536,r4
+noname.315:
+        movzwl  -26(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-28(fp),r0
+        ashl    #16,r0,-32(fp)
+        addl2   -32(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-32(fp)
+        bgequ   noname.316
+        incl    r4
+noname.316:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.317
+        incl    r2
+noname.317:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.318
+        incl    r9
+noname.318:
+
+        bicl3   #-65536,4(r6),r3
+        movzwl  6(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,4(r7),r2
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-36(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-40(fp)
+        mull2   r0,r4
+        addl3   -36(fp),-40(fp),r0
+        bicl3   #0,r0,-36(fp)
+        cmpl    -36(fp),-40(fp)
+        bgequ   noname.319
+        addl2   #65536,r4
+noname.319:
+        movzwl  -34(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-36(fp),r0
+        ashl    #16,r0,-40(fp)
+        addl2   -40(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-40(fp)
+        bgequ   noname.320
+        incl    r4
+noname.320:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.321
+        incl    r2
+noname.321:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.322
+        incl    r9
+noname.322:
+
+        bicl3   #-65536,(r6),r3
+        movzwl  2(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,8(r7),r2
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-44(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-48(fp)
+        mull2   r0,r4
+        addl3   -44(fp),-48(fp),r0
+        bicl3   #0,r0,-44(fp)
+        cmpl    -44(fp),-48(fp)
+        bgequ   noname.323
+        addl2   #65536,r4
+noname.323:
+        movzwl  -42(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-44(fp),r0
+        ashl    #16,r0,-48(fp)
+        addl2   -48(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-48(fp)
+        bgequ   noname.324
+        incl    r4
+noname.324:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.325
+        incl    r2
+noname.325:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.326
+        incl    r9
+noname.326:
+
+        movl    r8,8(r11)
+
+        clrl    r8
+
+        bicl3   #-65536,(r6),r3
+        movzwl  2(r6),r2
+        bicl3   #-65536,12(r7),r1
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r4
+        bicl3   #-65536,r2,-60(fp)
+        mull3   r0,r4,-52(fp)
+        mull2   r1,r4
+        mull3   r1,-60(fp),-56(fp)
+        mull2   r0,-60(fp)
+        addl3   -52(fp),-56(fp),r0
+        bicl3   #0,r0,-52(fp)
+        cmpl    -52(fp),-56(fp)
+        bgequ   noname.327
+        addl2   #65536,-60(fp)
+noname.327:
+        movzwl  -50(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-60(fp)
+        bicl3   #-65536,-52(fp),r0
+        ashl    #16,r0,-56(fp)
+        addl2   -56(fp),r4
+        bicl2   #0,r4
+        cmpl    r4,-56(fp)
+        bgequ   noname.328
+        incl    -60(fp)
+noname.328:
+        movl    r4,r1
+        movl    -60(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.329
+        incl    r2
+noname.329:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.330
+        incl    r8
+noname.330:
+
+        movzwl  6(r6),r2
+        bicl3   #-65536,8(r7),r3
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,4(r6),-72(fp)
+        bicl3   #-65536,r2,-76(fp)
+        mull3   r0,-72(fp),-64(fp)
+        mull2   r3,-72(fp)
+        mull3   r3,-76(fp),-68(fp)
+        mull2   r0,-76(fp)
+        addl3   -64(fp),-68(fp),r0
+        bicl3   #0,r0,-64(fp)
+        cmpl    -64(fp),-68(fp)
+        bgequ   noname.331
+        addl2   #65536,-76(fp)
+noname.331:
+        movzwl  -62(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-76(fp)
+        bicl3   #-65536,-64(fp),r0
+        ashl    #16,r0,-68(fp)
+        addl3   -68(fp),-72(fp),r0
+        bicl3   #0,r0,-72(fp)
+        cmpl    -72(fp),-68(fp)
+        bgequ   noname.332
+        incl    -76(fp)
+noname.332:
+        movl    -72(fp),r1
+        movl    -76(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.333
+        incl    r2
+noname.333:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.334
+        incl    r8
+noname.334:
+
+        bicl3   #-65536,8(r6),r3
+        movzwl  10(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,4(r7),r2
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-80(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-84(fp)
+        mull2   r0,r4
+        addl3   -80(fp),-84(fp),r0
+        bicl3   #0,r0,-80(fp)
+        cmpl    -80(fp),-84(fp)
+        bgequ   noname.335
+        addl2   #65536,r4
+noname.335:
+        movzwl  -78(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-80(fp),r0
+        ashl    #16,r0,-84(fp)
+        addl2   -84(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-84(fp)
+        bgequ   noname.336
+        incl    r4
+noname.336:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.337
+        incl    r2
+noname.337:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.338
+        incl    r8
+noname.338:
+
+        bicl3   #-65536,12(r6),r3
+        movzwl  14(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,(r7),r2
+        movzwl  2(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-88(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-92(fp)
+        mull2   r0,r4
+        addl3   -88(fp),-92(fp),r0
+        bicl3   #0,r0,-88(fp)
+        cmpl    -88(fp),-92(fp)
+        bgequ   noname.339
+        addl2   #65536,r4
+noname.339:
+        movzwl  -86(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-88(fp),r0
+        ashl    #16,r0,-92(fp)
+        addl2   -92(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-92(fp)
+        bgequ   noname.340
+        incl    r4
+noname.340:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.341
+        incl    r2
+noname.341:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.342
+        incl    r8
+noname.342:
+
+        movl    r10,12(r11)
+
+        clrl    r10
+
+        bicl3   #-65536,12(r6),r3
+        movzwl  14(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,4(r7),r2
+        movzwl  6(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-96(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-100(fp)
+        mull2   r0,r4
+        addl3   -96(fp),-100(fp),r0
+        bicl3   #0,r0,-96(fp)
+        cmpl    -96(fp),-100(fp)
+        bgequ   noname.343
+        addl2   #65536,r4
+noname.343:
+        movzwl  -94(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-96(fp),r0
+        ashl    #16,r0,-100(fp)
+        addl2   -100(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-100(fp)
+        bgequ   noname.344
+        incl    r4
+noname.344:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.345
+        incl    r2
+noname.345:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.346
+        incl    r10
+noname.346:
+
+        bicl3   #-65536,8(r6),r3
+        movzwl  10(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,8(r7),r2
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-104(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-108(fp)
+        mull2   r0,r4
+        addl3   -104(fp),-108(fp),r0
+        bicl3   #0,r0,-104(fp)
+        cmpl    -104(fp),-108(fp)
+        bgequ   noname.347
+        addl2   #65536,r4
+noname.347:
+        movzwl  -102(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-104(fp),r0
+        ashl    #16,r0,-108(fp)
+        addl2   -108(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-108(fp)
+        bgequ   noname.348
+        incl    r4
+noname.348:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.349
+        incl    r2
+noname.349:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.350
+        incl    r10
+noname.350:
+
+        bicl3   #-65536,4(r6),r3
+        movzwl  6(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,12(r7),r2
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-112(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-116(fp)
+        mull2   r0,r4
+        addl3   -112(fp),-116(fp),r0
+        bicl3   #0,r0,-112(fp)
+        cmpl    -112(fp),-116(fp)
+        bgequ   noname.351
+        addl2   #65536,r4
+noname.351:
+        movzwl  -110(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-112(fp),r0
+        ashl    #16,r0,-116(fp)
+        addl2   -116(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-116(fp)
+        bgequ   noname.352
+        incl    r4
+noname.352:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.353
+        incl    r2
+noname.353:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.354
+        incl    r10
+noname.354:
+
+        movl    r9,16(r11)
+
+        clrl    r9
+
+        bicl3   #-65536,8(r6),r3
+        movzwl  10(r6),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,12(r7),r2
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-120(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-124(fp)
+        mull2   r0,r4
+        addl3   -120(fp),-124(fp),r0
+        bicl3   #0,r0,-120(fp)
+        cmpl    -120(fp),-124(fp)
+        bgequ   noname.355
+        addl2   #65536,r4
+noname.355:
+        movzwl  -118(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-120(fp),r0
+        ashl    #16,r0,-124(fp)
+        addl2   -124(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-124(fp)
+        bgequ   noname.356
+        incl    r4
+noname.356:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.357
+        incl    r2
+noname.357:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.358
+        incl    r9
+noname.358:
+
+        movzwl  14(r6),r2
+        bicl3   #-65536,8(r7),r3
+        movzwl  10(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,12(r6),-136(fp)
+        bicl3   #-65536,r2,-140(fp)
+        mull3   r0,-136(fp),-128(fp)
+        mull2   r3,-136(fp)
+        mull3   r3,-140(fp),-132(fp)
+        mull2   r0,-140(fp)
+        addl3   -128(fp),-132(fp),r0
+        bicl3   #0,r0,-128(fp)
+        cmpl    -128(fp),-132(fp)
+        bgequ   noname.359
+        addl2   #65536,-140(fp)
+noname.359:
+        movzwl  -126(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-140(fp)
+        bicl3   #-65536,-128(fp),r0
+        ashl    #16,r0,-132(fp)
+        addl3   -132(fp),-136(fp),r0
+        bicl3   #0,r0,-136(fp)
+        cmpl    -136(fp),-132(fp)
+        bgequ   noname.360
+        incl    -140(fp)
+noname.360:
+        movl    -136(fp),r1
+        movl    -140(fp),r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.361
+        incl    r2
+noname.361:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.362
+        incl    r9
+noname.362:
+
+        movl    r8,20(r11)
+
+        clrl    r8
+
+        movzwl  14(r6),r2
+        bicl3   #-65536,12(r7),r3
+        movzwl  14(r7),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,12(r6),-152(fp)
+        bicl3   #-65536,r2,-156(fp)
+        mull3   r0,-152(fp),-144(fp)
+        mull2   r3,-152(fp)
+        mull3   r3,-156(fp),-148(fp)
+        mull2   r0,-156(fp)
+        addl3   -144(fp),-148(fp),r0
+        bicl3   #0,r0,-144(fp)
+        cmpl    -144(fp),-148(fp)
+        bgequ   noname.363
+        addl2   #65536,-156(fp)
+noname.363:
+        movzwl  -142(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-156(fp)
+        bicl3   #-65536,-144(fp),r0
+        ashl    #16,r0,-148(fp)
+        addl3   -148(fp),-152(fp),r0
+        bicl3   #0,r0,-152(fp)
+        cmpl    -152(fp),-148(fp)
+        bgequ   noname.364
+        incl    -156(fp)
+noname.364:
+        movl    -152(fp),r1
+        movl    -156(fp),r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.365
+        incl    r2
+noname.365:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.366
+        incl    r8
+noname.366:
+
+        movl    r10,24(r11)
+
+        movl    r9,28(r11)
+
+        ret     
+
+
+
+;r=4 ;(AP)
+;a=8 ;(AP)
+;b=12 ;(AP)
+;n=16 ;(AP)     n       by value (input)
+
+        .psect  code,nowrt
+
+.entry  BN_SQR_COMBA8,^m<r2,r3,r4,r5,r6,r7,r8,r9>
+        movab   -444(sp),sp
+
+        clrq    r8
+
+        clrl    r7
+
+        movl    8(ap),r4
+        movl    (r4),r3
+        bicl3   #-65536,r3,-4(fp)
+        extzv   #16,#16,r3,r0
+        bicl3   #-65536,r0,r3
+        movl    -4(fp),r0
+        mull3   r0,r3,-8(fp)
+        mull3   r0,r0,-4(fp)
+        mull2   r3,r3
+        bicl3   #32767,-8(fp),r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r3
+        bicl3   #-65536,-8(fp),r0
+        ashl    #17,r0,-8(fp)
+        addl3   -4(fp),-8(fp),r0
+        bicl3   #0,r0,-4(fp)
+        cmpl    -4(fp),-8(fp)
+        bgequ   noname.369
+        incl    r3
+noname.369:
+        movl    -4(fp),r1
+        movl    r3,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.370
+        incl    r2
+noname.370:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.371
+        incl    r7
+noname.371:
+
+        movl    r9,@4(ap)
+
+        clrl    r9
+
+        movzwl  6(r4),r2
+        bicl3   #-65536,(r4),r3
+        movzwl  2(r4),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,4(r4),-20(fp)
+        bicl3   #-65536,r2,-24(fp)
+        mull3   r0,-20(fp),-12(fp)
+        mull2   r3,-20(fp)
+        mull3   r3,-24(fp),-16(fp)
+        mull2   r0,-24(fp)
+        addl3   -12(fp),-16(fp),r0
+        bicl3   #0,r0,-12(fp)
+        cmpl    -12(fp),-16(fp)
+        bgequ   noname.372
+        addl2   #65536,-24(fp)
+noname.372:
+        movzwl  -10(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-24(fp)
+        bicl3   #-65536,-12(fp),r0
+        ashl    #16,r0,-16(fp)
+        addl3   -16(fp),-20(fp),r0
+        bicl3   #0,r0,-20(fp)
+        cmpl    -20(fp),-16(fp)
+        bgequ   noname.373
+        incl    -24(fp)
+noname.373:
+        movl    -20(fp),r3
+        movl    -24(fp),r2
+        bbc     #31,r2,noname.374
+        incl    r9
+noname.374:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.375
+        incl    r2
+noname.375:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.376
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.376
+        incl    r9
+noname.376:
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.377
+        incl    r9
+noname.377:
+
+        movl    4(ap),r0
+        movl    r8,4(r0)
+
+        clrl    r8
+
+        movl    8(ap),r4
+        movl    4(r4),r3
+        bicl3   #-65536,r3,-28(fp)
+        extzv   #16,#16,r3,r0
+        bicl3   #-65536,r0,r3
+        movl    -28(fp),r0
+        mull3   r0,r3,-32(fp)
+        mull3   r0,r0,-28(fp)
+        mull2   r3,r3
+        bicl3   #32767,-32(fp),r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r3
+        bicl3   #-65536,-32(fp),r0
+        ashl    #17,r0,-32(fp)
+        addl3   -28(fp),-32(fp),r0
+        bicl3   #0,r0,-28(fp)
+        cmpl    -28(fp),-32(fp)
+        bgequ   noname.378
+        incl    r3
+noname.378:
+        movl    -28(fp),r1
+        movl    r3,r2
+        addl2   r1,r7
+        bicl2   #0,r7
+        cmpl    r7,r1
+        bgequ   noname.379
+        incl    r2
+noname.379:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.380
+        incl    r8
+noname.380:
+
+        movzwl  10(r4),r2
+        bicl3   #-65536,(r4),r3
+        movzwl  2(r4),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,8(r4),-44(fp)
+        bicl3   #-65536,r2,-48(fp)
+        mull3   r0,-44(fp),-36(fp)
+        mull2   r3,-44(fp)
+        mull3   r3,-48(fp),-40(fp)
+        mull2   r0,-48(fp)
+        addl3   -36(fp),-40(fp),r0
+        bicl3   #0,r0,-36(fp)
+        cmpl    -36(fp),-40(fp)
+        bgequ   noname.381
+        addl2   #65536,-48(fp)
+noname.381:
+        movzwl  -34(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-48(fp)
+        bicl3   #-65536,-36(fp),r0
+        ashl    #16,r0,-40(fp)
+        addl3   -40(fp),-44(fp),r0
+        bicl3   #0,r0,-44(fp)
+        cmpl    -44(fp),-40(fp)
+        bgequ   noname.382
+        incl    -48(fp)
+noname.382:
+        movl    -44(fp),r3
+        movl    -48(fp),r2
+        bbc     #31,r2,noname.383
+        incl    r8
+noname.383:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.384
+        incl    r2
+noname.384:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r7
+        bicl2   #0,r7
+        cmpl    r7,r3
+        bgequ   noname.385
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.385
+        incl    r8
+noname.385:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.386
+        incl    r8
+noname.386:
+
+        movl    4(ap),r0
+        movl    r7,8(r0)
+
+        clrl    r7
+
+        movl    8(ap),r0
+        movzwl  14(r0),r2
+        bicl3   #-65536,(r0),r3
+        movzwl  2(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,12(r0),-60(fp)
+        bicl3   #-65536,r2,-64(fp)
+        mull3   r1,-60(fp),-52(fp)
+        mull2   r3,-60(fp)
+        mull3   r3,-64(fp),-56(fp)
+        mull2   r1,-64(fp)
+        addl3   -52(fp),-56(fp),r0
+        bicl3   #0,r0,-52(fp)
+        cmpl    -52(fp),-56(fp)
+        bgequ   noname.387
+        addl2   #65536,-64(fp)
+noname.387:
+        movzwl  -50(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-64(fp)
+        bicl3   #-65536,-52(fp),r0
+        ashl    #16,r0,-56(fp)
+        addl3   -56(fp),-60(fp),r0
+        bicl3   #0,r0,-60(fp)
+        cmpl    -60(fp),-56(fp)
+        bgequ   noname.388
+        incl    -64(fp)
+noname.388:
+        movl    -60(fp),r3
+        movl    -64(fp),r2
+        bbc     #31,r2,noname.389
+        incl    r7
+noname.389:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.390
+        incl    r2
+noname.390:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r9
+        bicl2   #0,r9
+        cmpl    r9,r3
+        bgequ   noname.391
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.391
+        incl    r7
+noname.391:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.392
+        incl    r7
+noname.392:
+
+        movl    8(ap),r0
+        movzwl  10(r0),r2
+        bicl3   #-65536,4(r0),r3
+        movzwl  6(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,8(r0),-76(fp)
+        bicl3   #-65536,r2,-80(fp)
+        mull3   r1,-76(fp),-68(fp)
+        mull2   r3,-76(fp)
+        mull3   r3,-80(fp),-72(fp)
+        mull2   r1,-80(fp)
+        addl3   -68(fp),-72(fp),r0
+        bicl3   #0,r0,-68(fp)
+        cmpl    -68(fp),-72(fp)
+        bgequ   noname.393
+        addl2   #65536,-80(fp)
+noname.393:
+        movzwl  -66(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-80(fp)
+        bicl3   #-65536,-68(fp),r0
+        ashl    #16,r0,-72(fp)
+        addl3   -72(fp),-76(fp),r0
+        bicl3   #0,r0,-76(fp)
+        cmpl    -76(fp),-72(fp)
+        bgequ   noname.394
+        incl    -80(fp)
+noname.394:
+        movl    -76(fp),r3
+        movl    -80(fp),r2
+        bbc     #31,r2,noname.395
+        incl    r7
+noname.395:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.396
+        incl    r2
+noname.396:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r9
+        bicl2   #0,r9
+        cmpl    r9,r3
+        bgequ   noname.397
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.397
+        incl    r7
+noname.397:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.398
+        incl    r7
+noname.398:
+
+        movl    4(ap),r0
+        movl    r9,12(r0)
+
+        clrl    r9
+
+        movl    8(ap),r2
+        movl    8(r2),r4
+        bicl3   #-65536,r4,-84(fp)
+        extzv   #16,#16,r4,r0
+        bicl3   #-65536,r0,r4
+        movl    -84(fp),r0
+        mull3   r0,r4,-88(fp)
+        mull3   r0,r0,-84(fp)
+        mull2   r4,r4
+        bicl3   #32767,-88(fp),r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r4
+        bicl3   #-65536,-88(fp),r0
+        ashl    #17,r0,-88(fp)
+        addl3   -84(fp),-88(fp),r0
+        bicl3   #0,r0,-84(fp)
+        cmpl    -84(fp),-88(fp)
+        bgequ   noname.399
+        incl    r4
+noname.399:
+        movl    -84(fp),r1
+        movl    r4,r3
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.400
+        incl    r3
+noname.400:
+        addl2   r3,r7
+        bicl2   #0,r7
+        cmpl    r7,r3
+        bgequ   noname.401
+        incl    r9
+noname.401:
+
+        movzwl  14(r2),r3
+        bicl3   #-65536,4(r2),r1
+        movzwl  6(r2),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,12(r2),-100(fp)
+        bicl3   #-65536,r3,-104(fp)
+        mull3   r0,-100(fp),-92(fp)
+        mull2   r1,-100(fp)
+        mull3   r1,-104(fp),-96(fp)
+        mull2   r0,-104(fp)
+        addl3   -92(fp),-96(fp),r0
+        bicl3   #0,r0,-92(fp)
+        cmpl    -92(fp),-96(fp)
+        bgequ   noname.402
+        addl2   #65536,-104(fp)
+noname.402:
+        movzwl  -90(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-104(fp)
+        bicl3   #-65536,-92(fp),r0
+        ashl    #16,r0,-96(fp)
+        addl3   -96(fp),-100(fp),r0
+        bicl3   #0,r0,-100(fp)
+        cmpl    -100(fp),-96(fp)
+        bgequ   noname.403
+        incl    -104(fp)
+noname.403:
+        movl    -100(fp),r3
+        movl    -104(fp),r2
+        bbc     #31,r2,noname.404
+        incl    r9
+noname.404:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.405
+        incl    r2
+noname.405:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.406
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.406
+        incl    r9
+noname.406:
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.407
+        incl    r9
+noname.407:
+
+        movl    8(ap),r0
+        movzwl  18(r0),r2
+        bicl3   #-65536,(r0),r3
+        movzwl  2(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,16(r0),-116(fp)
+        bicl3   #-65536,r2,-120(fp)
+        mull3   r1,-116(fp),-108(fp)
+        mull2   r3,-116(fp)
+        mull3   r3,-120(fp),-112(fp)
+        mull2   r1,-120(fp)
+        addl3   -108(fp),-112(fp),r0
+        bicl3   #0,r0,-108(fp)
+        cmpl    -108(fp),-112(fp)
+        bgequ   noname.408
+        addl2   #65536,-120(fp)
+noname.408:
+        movzwl  -106(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-120(fp)
+        bicl3   #-65536,-108(fp),r0
+        ashl    #16,r0,-112(fp)
+        addl3   -112(fp),-116(fp),r0
+        bicl3   #0,r0,-116(fp)
+        cmpl    -116(fp),-112(fp)
+        bgequ   noname.409
+        incl    -120(fp)
+noname.409:
+        movl    -116(fp),r3
+        movl    -120(fp),r2
+        bbc     #31,r2,noname.410
+        incl    r9
+noname.410:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.411
+        incl    r2
+noname.411:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.412
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.412
+        incl    r9
+noname.412:
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.413
+        incl    r9
+noname.413:
+
+        movl    4(ap),r0
+        movl    r8,16(r0)
+
+        clrl    r8
+
+        movl    8(ap),r0
+        movzwl  22(r0),r2
+        bicl3   #-65536,(r0),r3
+        movzwl  2(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,20(r0),-132(fp)
+        bicl3   #-65536,r2,-136(fp)
+        mull3   r1,-132(fp),-124(fp)
+        mull2   r3,-132(fp)
+        mull3   r3,-136(fp),-128(fp)
+        mull2   r1,-136(fp)
+        addl3   -124(fp),-128(fp),r0
+        bicl3   #0,r0,-124(fp)
+        cmpl    -124(fp),-128(fp)
+        bgequ   noname.414
+        addl2   #65536,-136(fp)
+noname.414:
+        movzwl  -122(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-136(fp)
+        bicl3   #-65536,-124(fp),r0
+        ashl    #16,r0,-128(fp)
+        addl3   -128(fp),-132(fp),r0
+        bicl3   #0,r0,-132(fp)
+        cmpl    -132(fp),-128(fp)
+        bgequ   noname.415
+        incl    -136(fp)
+noname.415:
+        movl    -132(fp),r3
+        movl    -136(fp),r2
+        bbc     #31,r2,noname.416
+        incl    r8
+noname.416:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.417
+        incl    r2
+noname.417:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r7
+        bicl2   #0,r7
+        cmpl    r7,r3
+        bgequ   noname.418
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.418
+        incl    r8
+noname.418:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.419
+        incl    r8
+noname.419:
+
+        movl    8(ap),r0
+        movzwl  18(r0),r2
+        bicl3   #-65536,4(r0),r3
+        movzwl  6(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,16(r0),-148(fp)
+        bicl3   #-65536,r2,-152(fp)
+        mull3   r1,-148(fp),-140(fp)
+        mull2   r3,-148(fp)
+        mull3   r3,-152(fp),-144(fp)
+        mull2   r1,-152(fp)
+        addl3   -140(fp),-144(fp),r0
+        bicl3   #0,r0,-140(fp)
+        cmpl    -140(fp),-144(fp)
+        bgequ   noname.420
+        addl2   #65536,-152(fp)
+noname.420:
+        movzwl  -138(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-152(fp)
+        bicl3   #-65536,-140(fp),r0
+        ashl    #16,r0,-144(fp)
+        addl3   -144(fp),-148(fp),r0
+        bicl3   #0,r0,-148(fp)
+        cmpl    -148(fp),-144(fp)
+        bgequ   noname.421
+        incl    -152(fp)
+noname.421:
+        movl    -148(fp),r3
+        movl    -152(fp),r2
+        bbc     #31,r2,noname.422
+        incl    r8
+noname.422:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.423
+        incl    r2
+noname.423:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r7
+        bicl2   #0,r7
+        cmpl    r7,r3
+        bgequ   noname.424
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.424
+        incl    r8
+noname.424:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.425
+        incl    r8
+noname.425:
+
+        movl    8(ap),r0
+        movzwl  14(r0),r2
+        bicl3   #-65536,8(r0),r3
+        movzwl  10(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,12(r0),-164(fp)
+        bicl3   #-65536,r2,-168(fp)
+        mull3   r1,-164(fp),-156(fp)
+        mull2   r3,-164(fp)
+        mull3   r3,-168(fp),-160(fp)
+        mull2   r1,-168(fp)
+        addl3   -156(fp),-160(fp),r0
+        bicl3   #0,r0,-156(fp)
+        cmpl    -156(fp),-160(fp)
+        bgequ   noname.426
+        addl2   #65536,-168(fp)
+noname.426:
+        movzwl  -154(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-168(fp)
+        bicl3   #-65536,-156(fp),r0
+        ashl    #16,r0,-160(fp)
+        addl3   -160(fp),-164(fp),r0
+        bicl3   #0,r0,-164(fp)
+        cmpl    -164(fp),-160(fp)
+        bgequ   noname.427
+        incl    -168(fp)
+noname.427:
+        movl    -164(fp),r3
+        movl    -168(fp),r2
+        bbc     #31,r2,noname.428
+        incl    r8
+noname.428:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.429
+        incl    r2
+noname.429:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r7
+        bicl2   #0,r7
+        cmpl    r7,r3
+        bgequ   noname.430
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.430
+        incl    r8
+noname.430:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.431
+        incl    r8
+noname.431:
+
+        movl    4(ap),r0
+        movl    r7,20(r0)
+
+        clrl    r7
+
+        movl    8(ap),r2
+        movl    12(r2),r4
+        bicl3   #-65536,r4,-172(fp)
+        extzv   #16,#16,r4,r0
+        bicl3   #-65536,r0,r4
+        movl    -172(fp),r0
+        mull3   r0,r4,-176(fp)
+        mull3   r0,r0,-172(fp)
+        mull2   r4,r4
+        bicl3   #32767,-176(fp),r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r4
+        bicl3   #-65536,-176(fp),r0
+        ashl    #17,r0,-176(fp)
+        addl3   -172(fp),-176(fp),r0
+        bicl3   #0,r0,-172(fp)
+        cmpl    -172(fp),-176(fp)
+        bgequ   noname.432
+        incl    r4
+noname.432:
+        movl    -172(fp),r1
+        movl    r4,r3
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.433
+        incl    r3
+noname.433:
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.434
+        incl    r7
+noname.434:
+
+        movzwl  18(r2),r3
+        bicl3   #-65536,8(r2),r1
+        movzwl  10(r2),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,16(r2),-188(fp)
+        bicl3   #-65536,r3,-192(fp)
+        mull3   r0,-188(fp),-180(fp)
+        mull2   r1,-188(fp)
+        mull3   r1,-192(fp),-184(fp)
+        mull2   r0,-192(fp)
+        addl3   -180(fp),-184(fp),r0
+        bicl3   #0,r0,-180(fp)
+        cmpl    -180(fp),-184(fp)
+        bgequ   noname.435
+        addl2   #65536,-192(fp)
+noname.435:
+        movzwl  -178(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-192(fp)
+        bicl3   #-65536,-180(fp),r0
+        ashl    #16,r0,-184(fp)
+        addl3   -184(fp),-188(fp),r0
+        bicl3   #0,r0,-188(fp)
+        cmpl    -188(fp),-184(fp)
+        bgequ   noname.436
+        incl    -192(fp)
+noname.436:
+        movl    -188(fp),r3
+        movl    -192(fp),r2
+        bbc     #31,r2,noname.437
+        incl    r7
+noname.437:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.438
+        incl    r2
+noname.438:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r9
+        bicl2   #0,r9
+        cmpl    r9,r3
+        bgequ   noname.439
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.439
+        incl    r7
+noname.439:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.440
+        incl    r7
+noname.440:
+
+        movl    8(ap),r0
+        movzwl  22(r0),r2
+        bicl3   #-65536,4(r0),r3
+        movzwl  6(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,20(r0),-204(fp)
+        bicl3   #-65536,r2,-208(fp)
+        mull3   r1,-204(fp),-196(fp)
+        mull2   r3,-204(fp)
+        mull3   r3,-208(fp),-200(fp)
+        mull2   r1,-208(fp)
+        addl3   -196(fp),-200(fp),r0
+        bicl3   #0,r0,-196(fp)
+        cmpl    -196(fp),-200(fp)
+        bgequ   noname.441
+        addl2   #65536,-208(fp)
+noname.441:
+        movzwl  -194(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-208(fp)
+        bicl3   #-65536,-196(fp),r0
+        ashl    #16,r0,-200(fp)
+        addl3   -200(fp),-204(fp),r0
+        bicl3   #0,r0,-204(fp)
+        cmpl    -204(fp),-200(fp)
+        bgequ   noname.442
+        incl    -208(fp)
+noname.442:
+        movl    -204(fp),r3
+        movl    -208(fp),r2
+        bbc     #31,r2,noname.443
+        incl    r7
+noname.443:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.444
+        incl    r2
+noname.444:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r9
+        bicl2   #0,r9
+        cmpl    r9,r3
+        bgequ   noname.445
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.445
+        incl    r7
+noname.445:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.446
+        incl    r7
+noname.446:
+
+        movl    8(ap),r0
+        movzwl  26(r0),r2
+        bicl3   #-65536,(r0),r3
+        movzwl  2(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,24(r0),-220(fp)
+        bicl3   #-65536,r2,-224(fp)
+        mull3   r1,-220(fp),-212(fp)
+        mull2   r3,-220(fp)
+        mull3   r3,-224(fp),-216(fp)
+        mull2   r1,-224(fp)
+        addl3   -212(fp),-216(fp),r0
+        bicl3   #0,r0,-212(fp)
+        cmpl    -212(fp),-216(fp)
+        bgequ   noname.447
+        addl2   #65536,-224(fp)
+noname.447:
+        movzwl  -210(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-224(fp)
+        bicl3   #-65536,-212(fp),r0
+        ashl    #16,r0,-216(fp)
+        addl3   -216(fp),-220(fp),r0
+        bicl3   #0,r0,-220(fp)
+        cmpl    -220(fp),-216(fp)
+        bgequ   noname.448
+        incl    -224(fp)
+noname.448:
+        movl    -220(fp),r3
+        movl    -224(fp),r2
+        bbc     #31,r2,noname.449
+        incl    r7
+noname.449:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.450
+        incl    r2
+noname.450:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r9
+        bicl2   #0,r9
+        cmpl    r9,r3
+        bgequ   noname.451
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.451
+        incl    r7
+noname.451:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.452
+        incl    r7
+noname.452:
+
+        movl    4(ap),r0
+        movl    r9,24(r0)
+
+        clrl    r9
+
+        movl    8(ap),r0
+        movzwl  30(r0),r2
+        bicl3   #-65536,(r0),r3
+        movzwl  2(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,28(r0),-236(fp)
+        bicl3   #-65536,r2,-240(fp)
+        mull3   r1,-236(fp),-228(fp)
+        mull2   r3,-236(fp)
+        mull3   r3,-240(fp),-232(fp)
+        mull2   r1,-240(fp)
+        addl3   -228(fp),-232(fp),r0
+        bicl3   #0,r0,-228(fp)
+        cmpl    -228(fp),-232(fp)
+        bgequ   noname.453
+        addl2   #65536,-240(fp)
+noname.453:
+        movzwl  -226(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-240(fp)
+        bicl3   #-65536,-228(fp),r0
+        ashl    #16,r0,-232(fp)
+        addl3   -232(fp),-236(fp),r0
+        bicl3   #0,r0,-236(fp)
+        cmpl    -236(fp),-232(fp)
+        bgequ   noname.454
+        incl    -240(fp)
+noname.454:
+        movl    -236(fp),r3
+        movl    -240(fp),r2
+        bbc     #31,r2,noname.455
+        incl    r9
+noname.455:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.456
+        incl    r2
+noname.456:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.457
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.457
+        incl    r9
+noname.457:
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.458
+        incl    r9
+noname.458:
+
+        movl    8(ap),r0
+        movzwl  26(r0),r2
+        bicl3   #-65536,4(r0),r3
+        movzwl  6(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,24(r0),-252(fp)
+        bicl3   #-65536,r2,-256(fp)
+        mull3   r1,-252(fp),-244(fp)
+        mull2   r3,-252(fp)
+        mull3   r3,-256(fp),-248(fp)
+        mull2   r1,-256(fp)
+        addl3   -244(fp),-248(fp),r0
+        bicl3   #0,r0,-244(fp)
+        cmpl    -244(fp),-248(fp)
+        bgequ   noname.459
+        addl2   #65536,-256(fp)
+noname.459:
+        movzwl  -242(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-256(fp)
+        bicl3   #-65536,-244(fp),r0
+        ashl    #16,r0,-248(fp)
+        addl3   -248(fp),-252(fp),r0
+        bicl3   #0,r0,-252(fp)
+        cmpl    -252(fp),-248(fp)
+        bgequ   noname.460
+        incl    -256(fp)
+noname.460:
+        movl    -252(fp),r3
+        movl    -256(fp),r2
+        bbc     #31,r2,noname.461
+        incl    r9
+noname.461:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.462
+        incl    r2
+noname.462:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.463
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.463
+        incl    r9
+noname.463:
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.464
+        incl    r9
+noname.464:
+
+        movl    8(ap),r0
+        movzwl  22(r0),r2
+        bicl3   #-65536,8(r0),r3
+        movzwl  10(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,20(r0),-268(fp)
+        bicl3   #-65536,r2,-272(fp)
+        mull3   r1,-268(fp),-260(fp)
+        mull2   r3,-268(fp)
+        mull3   r3,-272(fp),-264(fp)
+        mull2   r1,-272(fp)
+        addl3   -260(fp),-264(fp),r0
+        bicl3   #0,r0,-260(fp)
+        cmpl    -260(fp),-264(fp)
+        bgequ   noname.465
+        addl2   #65536,-272(fp)
+noname.465:
+        movzwl  -258(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-272(fp)
+        bicl3   #-65536,-260(fp),r0
+        ashl    #16,r0,-264(fp)
+        addl3   -264(fp),-268(fp),r0
+        bicl3   #0,r0,-268(fp)
+        cmpl    -268(fp),-264(fp)
+        bgequ   noname.466
+        incl    -272(fp)
+noname.466:
+        movl    -268(fp),r3
+        movl    -272(fp),r2
+        bbc     #31,r2,noname.467
+        incl    r9
+noname.467:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.468
+        incl    r2
+noname.468:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.469
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.469
+        incl    r9
+noname.469:
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.470
+        incl    r9
+noname.470:
+
+        movl    8(ap),r0
+        movzwl  18(r0),r2
+        bicl3   #-65536,12(r0),r3
+        movzwl  14(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,16(r0),-284(fp)
+        bicl3   #-65536,r2,-288(fp)
+        mull3   r1,-284(fp),-276(fp)
+        mull2   r3,-284(fp)
+        mull3   r3,-288(fp),-280(fp)
+        mull2   r1,-288(fp)
+        addl3   -276(fp),-280(fp),r0
+        bicl3   #0,r0,-276(fp)
+        cmpl    -276(fp),-280(fp)
+        bgequ   noname.471
+        addl2   #65536,-288(fp)
+noname.471:
+        movzwl  -274(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-288(fp)
+        bicl3   #-65536,-276(fp),r0
+        ashl    #16,r0,-280(fp)
+        addl3   -280(fp),-284(fp),r0
+        bicl3   #0,r0,-284(fp)
+        cmpl    -284(fp),-280(fp)
+        bgequ   noname.472
+        incl    -288(fp)
+noname.472:
+        movl    -284(fp),r3
+        movl    -288(fp),r2
+        bbc     #31,r2,noname.473
+        incl    r9
+noname.473:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.474
+        incl    r2
+noname.474:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.475
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.475
+        incl    r9
+noname.475:
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.476
+        incl    r9
+noname.476:
+
+        movl    4(ap),r0
+        movl    r8,28(r0)
+
+        clrl    r8
+
+        movl    8(ap),r3
+        movl    16(r3),r4
+        bicl3   #-65536,r4,r5
+        extzv   #16,#16,r4,r0
+        bicl3   #-65536,r0,r4
+        mull3   r5,r4,-292(fp)
+        mull2   r5,r5
+        mull2   r4,r4
+        bicl3   #32767,-292(fp),r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r4
+        bicl3   #-65536,-292(fp),r0
+        ashl    #17,r0,-292(fp)
+        addl2   -292(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-292(fp)
+        bgequ   noname.477
+        incl    r4
+noname.477:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r7
+        bicl2   #0,r7
+        cmpl    r7,r1
+        bgequ   noname.478
+        incl    r2
+noname.478:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.479
+        incl    r8
+noname.479:
+
+        bicl3   #-65536,20(r3),r4
+        movzwl  22(r3),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,12(r3),r2
+        movzwl  14(r3),r0
+        bicl2   #-65536,r0
+        movl    r4,r6
+        movl    r1,r5
+        mull3   r0,r6,-296(fp)
+        mull2   r2,r6
+        mull3   r2,r5,-300(fp)
+        mull2   r0,r5
+        addl3   -296(fp),-300(fp),r0
+        bicl3   #0,r0,-296(fp)
+        cmpl    -296(fp),-300(fp)
+        bgequ   noname.480
+        addl2   #65536,r5
+noname.480:
+        movzwl  -294(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r5
+        bicl3   #-65536,-296(fp),r0
+        ashl    #16,r0,-300(fp)
+        addl2   -300(fp),r6
+        bicl2   #0,r6
+        cmpl    r6,-300(fp)
+        bgequ   noname.481
+        incl    r5
+noname.481:
+        movl    r6,r3
+        movl    r5,r2
+        bbc     #31,r2,noname.482
+        incl    r8
+noname.482:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.483
+        incl    r2
+noname.483:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r7
+        bicl2   #0,r7
+        cmpl    r7,r3
+        bgequ   noname.484
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.484
+        incl    r8
+noname.484:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.485
+        incl    r8
+noname.485:
+
+        movl    8(ap),r0
+        bicl3   #-65536,24(r0),r3
+        movzwl  26(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,8(r0),r2
+        movzwl  10(r0),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-304(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-308(fp)
+        mull2   r0,r4
+        addl3   -304(fp),-308(fp),r0
+        bicl3   #0,r0,-304(fp)
+        cmpl    -304(fp),-308(fp)
+        bgequ   noname.486
+        addl2   #65536,r4
+noname.486:
+        movzwl  -302(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-304(fp),r0
+        ashl    #16,r0,-308(fp)
+        addl2   -308(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-308(fp)
+        bgequ   noname.487
+        incl    r4
+noname.487:
+        movl    r5,r3
+        movl    r4,r2
+        bbc     #31,r2,noname.488
+        incl    r8
+noname.488:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.489
+        incl    r2
+noname.489:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r7
+        bicl2   #0,r7
+        cmpl    r7,r3
+        bgequ   noname.490
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.490
+        incl    r8
+noname.490:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.491
+        incl    r8
+noname.491:
+
+        movl    8(ap),r0
+        bicl3   #-65536,28(r0),r3
+        movzwl  30(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,4(r0),r2
+        movzwl  6(r0),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-312(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-316(fp)
+        mull2   r0,r4
+        addl3   -312(fp),-316(fp),r0
+        bicl3   #0,r0,-312(fp)
+        cmpl    -312(fp),-316(fp)
+        bgequ   noname.492
+        addl2   #65536,r4
+noname.492:
+        movzwl  -310(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-312(fp),r0
+        ashl    #16,r0,-316(fp)
+        addl2   -316(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-316(fp)
+        bgequ   noname.493
+        incl    r4
+noname.493:
+        movl    r5,r3
+        movl    r4,r2
+        bbc     #31,r2,noname.494
+        incl    r8
+noname.494:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.495
+        incl    r2
+noname.495:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r7
+        bicl2   #0,r7
+        cmpl    r7,r3
+        bgequ   noname.496
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.496
+        incl    r8
+noname.496:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.497
+        incl    r8
+noname.497:
+
+        movl    4(ap),r0
+        movl    r7,32(r0)
+
+        clrl    r7
+
+        movl    8(ap),r0
+        bicl3   #-65536,28(r0),r3
+        movzwl  30(r0),r2
+        bicl3   #-65536,8(r0),r1
+        movzwl  10(r0),r0
+        bicl2   #-65536,r0
+        movl    r3,r4
+        bicl3   #-65536,r2,-328(fp)
+        mull3   r0,r4,-320(fp)
+        mull2   r1,r4
+        mull3   r1,-328(fp),-324(fp)
+        mull2   r0,-328(fp)
+        addl3   -320(fp),-324(fp),r0
+        bicl3   #0,r0,-320(fp)
+        cmpl    -320(fp),-324(fp)
+        bgequ   noname.498
+        addl2   #65536,-328(fp)
+noname.498:
+        movzwl  -318(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-328(fp)
+        bicl3   #-65536,-320(fp),r0
+        ashl    #16,r0,-324(fp)
+        addl2   -324(fp),r4
+        bicl2   #0,r4
+        cmpl    r4,-324(fp)
+        bgequ   noname.499
+        incl    -328(fp)
+noname.499:
+        movl    r4,r3
+        movl    -328(fp),r2
+        bbc     #31,r2,noname.500
+        incl    r7
+noname.500:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.501
+        incl    r2
+noname.501:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r9
+        bicl2   #0,r9
+        cmpl    r9,r3
+        bgequ   noname.502
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.502
+        incl    r7
+noname.502:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.503
+        incl    r7
+noname.503:
+
+        movl    8(ap),r0
+        movzwl  26(r0),r2
+        bicl3   #-65536,12(r0),r3
+        movzwl  14(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,24(r0),-340(fp)
+        bicl3   #-65536,r2,-344(fp)
+        mull3   r1,-340(fp),-332(fp)
+        mull2   r3,-340(fp)
+        mull3   r3,-344(fp),-336(fp)
+        mull2   r1,-344(fp)
+        addl3   -332(fp),-336(fp),r0
+        bicl3   #0,r0,-332(fp)
+        cmpl    -332(fp),-336(fp)
+        bgequ   noname.504
+        addl2   #65536,-344(fp)
+noname.504:
+        movzwl  -330(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-344(fp)
+        bicl3   #-65536,-332(fp),r0
+        ashl    #16,r0,-336(fp)
+        addl3   -336(fp),-340(fp),r0
+        bicl3   #0,r0,-340(fp)
+        cmpl    -340(fp),-336(fp)
+        bgequ   noname.505
+        incl    -344(fp)
+noname.505:
+        movl    -340(fp),r3
+        movl    -344(fp),r2
+        bbc     #31,r2,noname.506
+        incl    r7
+noname.506:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.507
+        incl    r2
+noname.507:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r9
+        bicl2   #0,r9
+        cmpl    r9,r3
+        bgequ   noname.508
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.508
+        incl    r7
+noname.508:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.509
+        incl    r7
+noname.509:
+
+        movl    8(ap),r0
+        movzwl  22(r0),r2
+        bicl3   #-65536,16(r0),r3
+        movzwl  18(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,20(r0),-356(fp)
+        bicl3   #-65536,r2,-360(fp)
+        mull3   r1,-356(fp),-348(fp)
+        mull2   r3,-356(fp)
+        mull3   r3,-360(fp),-352(fp)
+        mull2   r1,-360(fp)
+        addl3   -348(fp),-352(fp),r0
+        bicl3   #0,r0,-348(fp)
+        cmpl    -348(fp),-352(fp)
+        bgequ   noname.510
+        addl2   #65536,-360(fp)
+noname.510:
+        movzwl  -346(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-360(fp)
+        bicl3   #-65536,-348(fp),r0
+        ashl    #16,r0,-352(fp)
+        addl3   -352(fp),-356(fp),r0
+        bicl3   #0,r0,-356(fp)
+        cmpl    -356(fp),-352(fp)
+        bgequ   noname.511
+        incl    -360(fp)
+noname.511:
+        movl    -356(fp),r3
+        movl    -360(fp),r2
+        bbc     #31,r2,noname.512
+        incl    r7
+noname.512:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.513
+        incl    r2
+noname.513:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r9
+        bicl2   #0,r9
+        cmpl    r9,r3
+        bgequ   noname.514
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.514
+        incl    r7
+noname.514:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.515
+        incl    r7
+noname.515:
+
+        movl    4(ap),r0
+        movl    r9,36(r0)
+
+        clrl    r9
+
+        movl    8(ap),r3
+        movl    20(r3),r4
+        bicl3   #-65536,r4,-364(fp)
+        extzv   #16,#16,r4,r0
+        bicl3   #-65536,r0,r4
+        movl    -364(fp),r0
+        mull3   r0,r4,-368(fp)
+        mull3   r0,r0,-364(fp)
+        mull2   r4,r4
+        bicl3   #32767,-368(fp),r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r4
+        bicl3   #-65536,-368(fp),r0
+        ashl    #17,r0,-368(fp)
+        addl3   -364(fp),-368(fp),r0
+        bicl3   #0,r0,-364(fp)
+        cmpl    -364(fp),-368(fp)
+        bgequ   noname.516
+        incl    r4
+noname.516:
+        movl    -364(fp),r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.517
+        incl    r2
+noname.517:
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.518
+        incl    r9
+noname.518:
+
+        bicl3   #-65536,24(r3),r4
+        movzwl  26(r3),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,16(r3),r2
+        movzwl  18(r3),r0
+        bicl2   #-65536,r0
+        movl    r4,r6
+        movl    r1,r5
+        mull3   r0,r6,-372(fp)
+        mull2   r2,r6
+        mull3   r2,r5,-376(fp)
+        mull2   r0,r5
+        addl3   -372(fp),-376(fp),r0
+        bicl3   #0,r0,-372(fp)
+        cmpl    -372(fp),-376(fp)
+        bgequ   noname.519
+        addl2   #65536,r5
+noname.519:
+        movzwl  -370(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r5
+        bicl3   #-65536,-372(fp),r0
+        ashl    #16,r0,-376(fp)
+        addl2   -376(fp),r6
+        bicl2   #0,r6
+        cmpl    r6,-376(fp)
+        bgequ   noname.520
+        incl    r5
+noname.520:
+        movl    r6,r3
+        movl    r5,r2
+        bbc     #31,r2,noname.521
+        incl    r9
+noname.521:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.522
+        incl    r2
+noname.522:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.523
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.523
+        incl    r9
+noname.523:
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.524
+        incl    r9
+noname.524:
+
+        movl    8(ap),r0
+        bicl3   #-65536,28(r0),r3
+        movzwl  30(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,12(r0),r2
+        movzwl  14(r0),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-380(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-384(fp)
+        mull2   r0,r4
+        addl3   -380(fp),-384(fp),r0
+        bicl3   #0,r0,-380(fp)
+        cmpl    -380(fp),-384(fp)
+        bgequ   noname.525
+        addl2   #65536,r4
+noname.525:
+        movzwl  -378(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-380(fp),r0
+        ashl    #16,r0,-384(fp)
+        addl2   -384(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-384(fp)
+        bgequ   noname.526
+        incl    r4
+noname.526:
+        movl    r5,r3
+        movl    r4,r2
+        bbc     #31,r2,noname.527
+        incl    r9
+noname.527:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.528
+        incl    r2
+noname.528:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.529
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.529
+        incl    r9
+noname.529:
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.530
+        incl    r9
+noname.530:
+        movl    4(ap),r0
+        movl    r8,40(r0)
+
+        clrl    r8
+
+        movl    8(ap),r0
+        bicl3   #-65536,28(r0),r3
+        movzwl  30(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,16(r0),r2
+        movzwl  18(r0),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-388(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-392(fp)
+        mull2   r0,r4
+        addl3   -388(fp),-392(fp),r0
+        bicl3   #0,r0,-388(fp)
+        cmpl    -388(fp),-392(fp)
+        bgequ   noname.531
+        addl2   #65536,r4
+noname.531:
+        movzwl  -386(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-388(fp),r0
+        ashl    #16,r0,-392(fp)
+        addl2   -392(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-392(fp)
+        bgequ   noname.532
+        incl    r4
+noname.532:
+        movl    r5,r3
+        movl    r4,r2
+        bbc     #31,r2,noname.533
+        incl    r8
+noname.533:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.534
+        incl    r2
+noname.534:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r7
+        bicl2   #0,r7
+        cmpl    r7,r3
+        bgequ   noname.535
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.535
+        incl    r8
+noname.535:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.536
+        incl    r8
+noname.536:
+
+        movl    8(ap),r0
+        bicl3   #-65536,24(r0),r3
+        movzwl  26(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,20(r0),r2
+        movzwl  22(r0),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-396(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-400(fp)
+        mull2   r0,r4
+        addl3   -396(fp),-400(fp),r0
+        bicl3   #0,r0,-396(fp)
+        cmpl    -396(fp),-400(fp)
+        bgequ   noname.537
+        addl2   #65536,r4
+noname.537:
+        movzwl  -394(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-396(fp),r0
+        ashl    #16,r0,-400(fp)
+        addl2   -400(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-400(fp)
+        bgequ   noname.538
+        incl    r4
+noname.538:
+        movl    r5,r3
+        movl    r4,r2
+        bbc     #31,r2,noname.539
+        incl    r8
+noname.539:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.540
+        incl    r2
+noname.540:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r7
+        bicl2   #0,r7
+        cmpl    r7,r3
+        bgequ   noname.541
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.541
+        incl    r8
+noname.541:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.542
+        incl    r8
+noname.542:
+
+        movl    4(ap),r0
+        movl    r7,44(r0)
+
+        clrl    r7
+
+        movl    8(ap),r3
+        movl    24(r3),r4
+        bicl3   #-65536,r4,r5
+        extzv   #16,#16,r4,r0
+        bicl3   #-65536,r0,r4
+        mull3   r5,r4,-404(fp)
+        mull2   r5,r5
+        mull2   r4,r4
+        bicl3   #32767,-404(fp),r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r4
+        bicl3   #-65536,-404(fp),r0
+        ashl    #17,r0,-404(fp)
+        addl2   -404(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-404(fp)
+        bgequ   noname.543
+        incl    r4
+noname.543:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.544
+        incl    r2
+noname.544:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.545
+        incl    r7
+noname.545:
+
+        movzwl  30(r3),r2
+        bicl3   #-65536,20(r3),r1
+        movzwl  22(r3),r0
+        bicl2   #-65536,r0
+        bicl3   #-65536,28(r3),-416(fp)
+        bicl3   #-65536,r2,-420(fp)
+        mull3   r0,-416(fp),-408(fp)
+        mull2   r1,-416(fp)
+        mull3   r1,-420(fp),-412(fp)
+        mull2   r0,-420(fp)
+        addl3   -408(fp),-412(fp),r0
+        bicl3   #0,r0,-408(fp)
+        cmpl    -408(fp),-412(fp)
+        bgequ   noname.546
+        addl2   #65536,-420(fp)
+noname.546:
+        movzwl  -406(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-420(fp)
+        bicl3   #-65536,-408(fp),r0
+        ashl    #16,r0,-412(fp)
+        addl3   -412(fp),-416(fp),r0
+        bicl3   #0,r0,-416(fp)
+        cmpl    -416(fp),-412(fp)
+        bgequ   noname.547
+        incl    -420(fp)
+noname.547:
+        movl    -416(fp),r3
+        movl    -420(fp),r2
+        bbc     #31,r2,noname.548
+        incl    r7
+noname.548:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.549
+        incl    r2
+noname.549:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r9
+        bicl2   #0,r9
+        cmpl    r9,r3
+        bgequ   noname.550
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.550
+        incl    r7
+noname.550:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.551
+        incl    r7
+noname.551:
+
+        movl    4(ap),r0
+        movl    r9,48(r0)
+
+        clrl    r9
+
+        movl    8(ap),r0
+        movzwl  30(r0),r2
+        bicl3   #-65536,24(r0),r3
+        movzwl  26(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,28(r0),-432(fp)
+        bicl3   #-65536,r2,-436(fp)
+        mull3   r1,-432(fp),-424(fp)
+        mull2   r3,-432(fp)
+        mull3   r3,-436(fp),-428(fp)
+        mull2   r1,-436(fp)
+        addl3   -424(fp),-428(fp),r0
+        bicl3   #0,r0,-424(fp)
+        cmpl    -424(fp),-428(fp)
+        bgequ   noname.552
+        addl2   #65536,-436(fp)
+noname.552:
+        movzwl  -422(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,-436(fp)
+        bicl3   #-65536,-424(fp),r0
+        ashl    #16,r0,-428(fp)
+        addl3   -428(fp),-432(fp),r0
+        bicl3   #0,r0,-432(fp)
+        cmpl    -432(fp),-428(fp)
+        bgequ   noname.553
+        incl    -436(fp)
+noname.553:
+        movl    -432(fp),r3
+        movl    -436(fp),r2
+        bbc     #31,r2,noname.554
+        incl    r9
+noname.554:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.555
+        incl    r2
+noname.555:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.556
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.556
+        incl    r9
+noname.556:
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.557
+        incl    r9
+noname.557:
+
+        movl    4(ap),r4
+        movl    r8,52(r4)
+
+        clrl    r8
+
+        movl    8(ap),r0
+        movl    28(r0),r3
+        bicl3   #-65536,r3,-440(fp)
+        extzv   #16,#16,r3,r0
+        bicl3   #-65536,r0,r3
+        movl    -440(fp),r0
+        mull3   r0,r3,-444(fp)
+        mull3   r0,r0,-440(fp)
+        mull2   r3,r3
+        bicl3   #32767,-444(fp),r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r3
+        bicl3   #-65536,-444(fp),r0
+        ashl    #17,r0,-444(fp)
+        addl3   -440(fp),-444(fp),r0
+        bicl3   #0,r0,-440(fp)
+        cmpl    -440(fp),-444(fp)
+        bgequ   noname.558
+        incl    r3
+noname.558:
+        movl    -440(fp),r1
+        movl    r3,r2
+        addl2   r1,r7
+        bicl2   #0,r7
+        cmpl    r7,r1
+        bgequ   noname.559
+        incl    r2
+noname.559:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.560
+        incl    r8
+noname.560:
+
+        movl    r7,56(r4)
+
+        movl    r9,60(r4)
+
+        ret     
+
+
+
+;r=4 ;(AP)
+;a=8 ;(AP)
+;b=12 ;(AP)
+;n=16 ;(AP)     n       by value (input)
+
+        .psect  code,nowrt
+
+.entry  BN_SQR_COMBA4,^m<r2,r3,r4,r5,r6,r7,r8,r9,r10>
+        subl2   #44,sp
+
+        clrq    r8
+
+        clrl    r10
+
+        movl    8(ap),r5
+        movl    (r5),r3
+        bicl3   #-65536,r3,r4
+        extzv   #16,#16,r3,r0
+        bicl3   #-65536,r0,r3
+        mull3   r4,r3,-4(fp)
+        mull2   r4,r4
+        mull2   r3,r3
+        bicl3   #32767,-4(fp),r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r3
+        bicl3   #-65536,-4(fp),r0
+        ashl    #17,r0,-4(fp)
+        addl2   -4(fp),r4
+        bicl2   #0,r4
+        cmpl    r4,-4(fp)
+        bgequ   noname.563
+        incl    r3
+noname.563:
+        movl    r4,r1
+        movl    r3,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.564
+        incl    r2
+noname.564:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.565
+        incl    r10
+noname.565:
+
+        movl    r9,@4(ap)
+
+        clrl    r9
+
+        bicl3   #-65536,4(r5),r3
+        movzwl  6(r5),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,(r5),r2
+        movzwl  2(r5),r0
+        bicl2   #-65536,r0
+        movl    r3,r6
+        movl    r1,r4
+        mull3   r0,r6,-8(fp)
+        mull2   r2,r6
+        mull2   r4,r2
+        mull2   r0,r4
+        addl3   -8(fp),r2,r0
+        bicl3   #0,r0,-8(fp)
+        cmpl    -8(fp),r2
+        bgequ   noname.566
+        addl2   #65536,r4
+noname.566:
+        movzwl  -6(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-8(fp),r0
+        ashl    #16,r0,r1
+        addl2   r1,r6
+        bicl2   #0,r6
+        cmpl    r6,r1
+        bgequ   noname.567
+        incl    r4
+noname.567:
+        movl    r6,r3
+        movl    r4,r2
+        bbc     #31,r2,noname.568
+        incl    r9
+noname.568:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.569
+        incl    r2
+noname.569:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.570
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.570
+        incl    r9
+noname.570:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.571
+        incl    r9
+noname.571:
+
+        movl    4(ap),r0
+        movl    r8,4(r0)
+
+        clrl    r8
+
+        movl    8(ap),r4
+        movl    4(r4),r3
+        bicl3   #-65536,r3,r5
+        extzv   #16,#16,r3,r0
+        bicl3   #-65536,r0,r3
+        mull3   r5,r3,r1
+        mull2   r5,r5
+        mull2   r3,r3
+        bicl3   #32767,r1,r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r3
+        bicl2   #-65536,r1
+        ashl    #17,r1,r1
+        addl2   r1,r5
+        bicl2   #0,r5
+        cmpl    r5,r1
+        bgequ   noname.572
+        incl    r3
+noname.572:
+        movl    r5,r1
+        movl    r3,r2
+        addl2   r1,r10
+        bicl2   #0,r10
+        cmpl    r10,r1
+        bgequ   noname.573
+        incl    r2
+noname.573:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.574
+        incl    r8
+noname.574:
+
+        bicl3   #-65536,8(r4),r3
+        movzwl  10(r4),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,(r4),r2
+        movzwl  2(r4),r0
+        bicl2   #-65536,r0
+        movl    r3,r6
+        movl    r1,r5
+        mull3   r0,r6,r7
+        mull2   r2,r6
+        mull2   r5,r2
+        mull2   r0,r5
+        addl2   r2,r7
+        bicl2   #0,r7
+        cmpl    r7,r2
+        bgequ   noname.575
+        addl2   #65536,r5
+noname.575:
+        extzv   #16,#16,r7,r0
+        bicl2   #-65536,r0
+        addl2   r0,r5
+        bicl3   #-65536,r7,r0
+        ashl    #16,r0,r1
+        addl2   r1,r6
+        bicl2   #0,r6
+        cmpl    r6,r1
+        bgequ   noname.576
+        incl    r5
+noname.576:
+        movl    r6,r3
+        movl    r5,r2
+        bbc     #31,r2,noname.577
+        incl    r8
+noname.577:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.578
+        incl    r2
+noname.578:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r10
+        bicl2   #0,r10
+        cmpl    r10,r3
+        bgequ   noname.579
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.579
+        incl    r8
+noname.579:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.580
+        incl    r8
+noname.580:
+
+        movl    4(ap),r0
+        movl    r10,8(r0)
+
+        clrl    r10
+
+        movl    8(ap),r0
+        bicl3   #-65536,12(r0),r3
+        movzwl  14(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,(r0),r2
+        movzwl  2(r0),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,r6
+        mull2   r2,r5
+        mull3   r2,r4,-12(fp)
+        mull2   r0,r4
+        addl2   -12(fp),r6
+        bicl2   #0,r6
+        cmpl    r6,-12(fp)
+        bgequ   noname.581
+        addl2   #65536,r4
+noname.581:
+        extzv   #16,#16,r6,r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,r6,r0
+        ashl    #16,r0,-12(fp)
+        addl2   -12(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-12(fp)
+        bgequ   noname.582
+        incl    r4
+noname.582:
+        movl    r5,r3
+        movl    r4,r2
+        bbc     #31,r2,noname.583
+        incl    r10
+noname.583:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.584
+        incl    r2
+noname.584:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r9
+        bicl2   #0,r9
+        cmpl    r9,r3
+        bgequ   noname.585
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.585
+        incl    r10
+noname.585:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.586
+        incl    r10
+noname.586:
+
+        movl    8(ap),r0
+        bicl3   #-65536,8(r0),r3
+        movzwl  10(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,4(r0),r2
+        movzwl  6(r0),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-16(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-20(fp)
+        mull2   r0,r4
+        addl3   -16(fp),-20(fp),r0
+        bicl3   #0,r0,-16(fp)
+        cmpl    -16(fp),-20(fp)
+        bgequ   noname.587
+        addl2   #65536,r4
+noname.587:
+        movzwl  -14(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-16(fp),r0
+        ashl    #16,r0,-20(fp)
+        addl2   -20(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-20(fp)
+        bgequ   noname.588
+        incl    r4
+noname.588:
+        movl    r5,r3
+        movl    r4,r2
+        bbc     #31,r2,noname.589
+        incl    r10
+noname.589:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.590
+        incl    r2
+noname.590:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r9
+        bicl2   #0,r9
+        cmpl    r9,r3
+        bgequ   noname.591
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.591
+        incl    r10
+noname.591:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.592
+        incl    r10
+noname.592:
+        movl    4(ap),r0
+        movl    r9,12(r0)
+
+        clrl    r9
+
+        movl    8(ap),r3
+        movl    8(r3),r4
+        bicl3   #-65536,r4,r5
+        extzv   #16,#16,r4,r0
+        bicl3   #-65536,r0,r4
+        mull3   r5,r4,-24(fp)
+        mull2   r5,r5
+        mull2   r4,r4
+        bicl3   #32767,-24(fp),r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r4
+        bicl3   #-65536,-24(fp),r0
+        ashl    #17,r0,-24(fp)
+        addl2   -24(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-24(fp)
+        bgequ   noname.593
+        incl    r4
+noname.593:
+        movl    r5,r1
+        movl    r4,r2
+        addl2   r1,r8
+        bicl2   #0,r8
+        cmpl    r8,r1
+        bgequ   noname.594
+        incl    r2
+noname.594:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.595
+        incl    r9
+noname.595:
+
+        bicl3   #-65536,12(r3),r4
+        movzwl  14(r3),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,4(r3),r2
+        movzwl  6(r3),r0
+        bicl2   #-65536,r0
+        movl    r4,r6
+        movl    r1,r5
+        mull3   r0,r6,-28(fp)
+        mull2   r2,r6
+        mull3   r2,r5,-32(fp)
+        mull2   r0,r5
+        addl3   -28(fp),-32(fp),r0
+        bicl3   #0,r0,-28(fp)
+        cmpl    -28(fp),-32(fp)
+        bgequ   noname.596
+        addl2   #65536,r5
+noname.596:
+        movzwl  -26(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r5
+        bicl3   #-65536,-28(fp),r0
+        ashl    #16,r0,-32(fp)
+        addl2   -32(fp),r6
+        bicl2   #0,r6
+        cmpl    r6,-32(fp)
+        bgequ   noname.597
+        incl    r5
+noname.597:
+        movl    r6,r3
+        movl    r5,r2
+        bbc     #31,r2,noname.598
+        incl    r9
+noname.598:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.599
+        incl    r2
+noname.599:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r8
+        bicl2   #0,r8
+        cmpl    r8,r3
+        bgequ   noname.600
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.600
+        incl    r9
+noname.600:
+        addl2   r2,r10
+        bicl2   #0,r10
+        cmpl    r10,r2
+        bgequ   noname.601
+        incl    r9
+noname.601:
+
+        movl    4(ap),r0
+        movl    r8,16(r0)
+
+        clrl    r8
+
+        movl    8(ap),r0
+        bicl3   #-65536,12(r0),r3
+        movzwl  14(r0),r1
+        bicl2   #-65536,r1
+        bicl3   #-65536,8(r0),r2
+        movzwl  10(r0),r0
+        bicl2   #-65536,r0
+        movl    r3,r5
+        movl    r1,r4
+        mull3   r0,r5,-36(fp)
+        mull2   r2,r5
+        mull3   r2,r4,-40(fp)
+        mull2   r0,r4
+        addl3   -36(fp),-40(fp),r0
+        bicl3   #0,r0,-36(fp)
+        cmpl    -36(fp),-40(fp)
+        bgequ   noname.602
+        addl2   #65536,r4
+noname.602:
+        movzwl  -34(fp),r0
+        bicl2   #-65536,r0
+        addl2   r0,r4
+        bicl3   #-65536,-36(fp),r0
+        ashl    #16,r0,-40(fp)
+        addl2   -40(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-40(fp)
+        bgequ   noname.603
+        incl    r4
+noname.603:
+        movl    r5,r3
+        movl    r4,r2
+        bbc     #31,r2,noname.604
+        incl    r8
+noname.604:
+        addl2   r2,r2
+        bicl2   #0,r2
+        bbc     #31,r3,noname.605
+        incl    r2
+noname.605:
+        addl2   r3,r3
+        bicl2   #0,r3
+        addl2   r3,r10
+        bicl2   #0,r10
+        cmpl    r10,r3
+        bgequ   noname.606
+        incl    r2
+        bicl3   #0,r2,r0
+        bneq    noname.606
+        incl    r8
+noname.606:
+        addl2   r2,r9
+        bicl2   #0,r9
+        cmpl    r9,r2
+        bgequ   noname.607
+        incl    r8
+noname.607:
+
+        movl    4(ap),r4
+        movl    r10,20(r4)
+
+        clrl    r10
+
+        movl    8(ap),r0
+        movl    12(r0),r3
+        bicl3   #-65536,r3,r5
+        extzv   #16,#16,r3,r0
+        bicl3   #-65536,r0,r3
+        mull3   r5,r3,-44(fp)
+        mull2   r5,r5
+        mull2   r3,r3
+        bicl3   #32767,-44(fp),r0
+        extzv   #15,#17,r0,r0
+        addl2   r0,r3
+        bicl3   #-65536,-44(fp),r0
+        ashl    #17,r0,-44(fp)
+        addl2   -44(fp),r5
+        bicl2   #0,r5
+        cmpl    r5,-44(fp)
+        bgequ   noname.608
+        incl    r3
+noname.608:
+        movl    r5,r1
+        movl    r3,r2
+        addl2   r1,r9
+        bicl2   #0,r9
+        cmpl    r9,r1
+        bgequ   noname.609
+        incl    r2
+noname.609:
+        addl2   r2,r8
+        bicl2   #0,r8
+        cmpl    r8,r2
+        bgequ   noname.610
+        incl    r10
+noname.610:
+
+        movl    r9,24(r4)
+
+        movl    r8,28(r4)
+
+        ret     
+
+; 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/asm/x86.pl b/src/lib/libcrypto/bn/asm/x86.pl
new file mode 100644
index 0000000000..1bc4f1bb27
--- /dev/null
+++ b/src/lib/libcrypto/bn/asm/x86.pl
@@ -0,0 +1,28 @@
+#!/usr/local/bin/perl
+
+push(@INC,"perlasm","../../perlasm");
+require "x86asm.pl";
+
+require("x86/mul_add.pl");
+require("x86/mul.pl");
+require("x86/sqr.pl");
+require("x86/div.pl");
+require("x86/add.pl");
+require("x86/sub.pl");
+require("x86/comba.pl");
+
+&asm_init($ARGV[0],$0);
+
+&bn_mul_add_words("bn_mul_add_words");
+&bn_mul_words("bn_mul_words");
+&bn_sqr_words("bn_sqr_words");
+&bn_div_words("bn_div_words");
+&bn_add_words("bn_add_words");
+&bn_sub_words("bn_sub_words");
+&bn_mul_comba("bn_mul_comba8",8);
+&bn_mul_comba("bn_mul_comba4",4);
+&bn_sqr_comba("bn_sqr_comba8",8);
+&bn_sqr_comba("bn_sqr_comba4",4);
+
+&asm_finish();
+
diff --git a/src/lib/libcrypto/bn/asm/x86w16.asm b/src/lib/libcrypto/bn/asm/x86w16.asm
index 74a933a8cd..80a9ed6eef 100644
--- a/src/lib/libcrypto/bn/asm/x86w16.asm
+++ b/src/lib/libcrypto/bn/asm/x86w16.asm
@@ -6,11 +6,11 @@ F_TEXT	SEGMENT  WORD PUBLIC 'CODE'
 F_TEXT  ENDS
 _DATA   SEGMENT  WORD PUBLIC 'DATA'
 _DATA   ENDS
-CONST   SEGMENT  WORD PUBLIC 'CONST'
-CONST   ENDS
+_CONST  SEGMENT  WORD PUBLIC 'CONST'
+_CONST  ENDS
 _BSS    SEGMENT  WORD PUBLIC 'BSS'
 _BSS    ENDS
-DGROUP  GROUP   CONST, _BSS, _DATA
+DGROUP  GROUP   _CONST, _BSS, _DATA
         ASSUME DS: DGROUP, SS: DGROUP
 F_TEXT      SEGMENT
         ASSUME  CS: F_TEXT
diff --git a/src/lib/libcrypto/bn/asm/x86w32.asm b/src/lib/libcrypto/bn/asm/x86w32.asm
index fc6f917714..957d71e3b1 100644
--- a/src/lib/libcrypto/bn/asm/x86w32.asm
+++ b/src/lib/libcrypto/bn/asm/x86w32.asm
@@ -6,11 +6,11 @@ F_TEXT	SEGMENT  WORD USE16 PUBLIC 'CODE'
 F_TEXT  ENDS
 _DATA   SEGMENT  WORD USE16 PUBLIC 'DATA'
 _DATA   ENDS
-CONST   SEGMENT  WORD USE16 PUBLIC 'CONST'
-CONST   ENDS
+_CONST  SEGMENT  WORD USE16 PUBLIC 'CONST'
+_CONST  ENDS
 _BSS    SEGMENT  WORD USE16 PUBLIC 'BSS'
 _BSS    ENDS
-DGROUP  GROUP   CONST, _BSS, _DATA
+DGROUP  GROUP   _CONST, _BSS, _DATA
         ASSUME DS: DGROUP, SS: DGROUP
 F_TEXT      SEGMENT
         ASSUME  CS: F_TEXT
@@ -89,7 +89,7 @@ $L555:
         mov     bp,WORD PTR [bp+26]     ; load num
         and     bp,3
         dec     bp
-        js      $L547
+        js      $L547m
 
         mov     eax,ecx
         mul     DWORD PTR es:[bx]       ; w* *a
@@ -100,7 +100,7 @@ $L555:
         mov     DWORD PTR ds:[di],eax
         mov     esi,edx
         dec     bp
-        js      $L547                   ; Note that we are now testing for -1
+        js      $L547m                  ; Note that we are now testing for -1
         ;
         mov     eax,ecx
         mul     DWORD PTR es:[bx+4]     ; w* *a
@@ -111,7 +111,7 @@ $L555:
         mov     DWORD PTR ds:[di+4],eax
         mov     esi,edx
         dec     bp
-        js      $L547
+        js      $L547m
         ;
         mov     eax,ecx
         mul     DWORD PTR es:[bx+8]     ; w* *a
@@ -121,7 +121,7 @@ $L555:
         adc     edx,0
         mov     DWORD PTR ds:[di+8],eax
         mov     esi,edx
-$L547:
+$L547m:
         mov     eax,esi
         mov     edx,esi
         shr     edx,16
@@ -315,37 +315,35 @@ _bn_add_words	PROC FAR
 ;       ap = 22
 ;       rp = 18
         xor     esi,esi                 ;c=0;
+        mov     bx,WORD PTR [bp+18]     ; load low r
         mov     si,WORD PTR [bp+22]     ; load a
         mov     es,WORD PTR [bp+24]     ; load a
         mov     di,WORD PTR [bp+26]     ; load b
         mov     ds,WORD PTR [bp+28]     ; load b
 
         mov     dx,WORD PTR [bp+30]     ; load num
-        dec     dx
-        js      $L547
         xor     ecx,ecx
+        dec     dx
+        js      $L547a
 
 $L5477:
-        xor     ebx,ebx
         mov     eax,DWORD PTR es:[si]   ; *a
         add     eax,ecx
-        adc     ebx,0
+        mov     ecx,0
+        adc     ecx,0
         add     si,4                    ; a++
         add     eax,DWORD PTR ds:[di]   ; + *b
-        mov     ecx,ebx
         adc     ecx,0
-        add     di,4
-        mov     bx,WORD PTR [bp+18]
         mov     ds,WORD PTR [bp+20]
+        add     di,4
         mov     DWORD PTR ds:[bx],eax
-        add     bx,4
         mov     ds,WORD PTR [bp+28]
-        mov     WORD PTR [bp+18],bx
+        add     bx,4
         dec     dx
-        js      $L547                   ; Note that we are now testing for -1
+        js      $L547a                  ; Note that we are now testing for -1
         jmp     $L5477
         ;
-$L547:
+$L547a:
         mov     eax,ecx
         mov     edx,ecx
         shr     edx,16