Implement decryption in the Serpent assembly code

4 files changed, 386 insertions, 207 deletions

diff --git a/modules/alg_ia32/modinfo.txt b/modules/alg_ia32/modinfo.txt
index 02581578f..1f85f5497 100644
--- a/modules/alg_ia32/modinfo.txt
+++ b/modules/alg_ia32/modinfo.txt
@@ -4,6 +4,7 @@ replace_file md4.cpp
 replace_file md5.cpp
 replace_file sha160.cpp
 replace_file serpent.cpp
+replace_file serpent.h
 
 add_file asm_macr.h
 
diff --git a/modules/alg_ia32/serp_asm.S b/modules/alg_ia32/serp_asm.S
index 049c2976b..a2785d129 100644
--- a/modules/alg_ia32/serp_asm.S
+++ b/modules/alg_ia32/serp_asm.S
@@ -7,7 +7,7 @@
 
 START_LISTING(serp_asm.S)
 
-#define SBOX1(A, B, C, D, T)       \
+#define SBOX_E1(A, B, C, D, T)     \
    XOR(D, A)                     ; \
    ASSIGN(T, B)                  ; \
    AND(B, D)                     ; \
@@ -30,7 +30,7 @@ START_LISTING(serp_asm.S)
    ASSIGN(A, B)                  ; \
    ASSIGN(B, T)                  ;
 
-#define SBOX2(A, B, C, D, T)       \
+#define SBOX_E2(A, B, C, D, T)     \
    NOT(A)                        ; \
    NOT(C)                        ; \
    ASSIGN(T, A)                  ; \
@@ -54,7 +54,7 @@ START_LISTING(serp_asm.S)
    ASSIGN(D, B)                  ; \
    ASSIGN(B, T)                  ;
 
-#define SBOX3(A, B, C, D, T)       \
+#define SBOX_E3(A, B, C, D, T)     \
    ASSIGN(T, A)                  ; \
    AND(A, C)                     ; \
    XOR(A, D)                     ; \
@@ -76,7 +76,7 @@ START_LISTING(serp_asm.S)
    ASSIGN(B, D)                  ; \
    ASSIGN(D, T)                  ;
 
-#define SBOX4(A, B, C, D, T)       \
+#define SBOX_E4(A, B, C, D, T)     \
    ASSIGN(T, A)                  ; \
    OR(A, D)                      ; \
    XOR(D, B)                     ; \
@@ -101,7 +101,7 @@ START_LISTING(serp_asm.S)
    ASSIGN(C, D)                  ; \
    ASSIGN(D, T)                  ; \
 
-#define SBOX5(A, B, C, D, T)       \
+#define SBOX_E5(A, B, C, D, T)     \
    XOR(B, D)                     ; \
    NOT(D)                        ; \
    XOR(C, D)                     ; \
@@ -126,7 +126,7 @@ START_LISTING(serp_asm.S)
    ASSIGN(A, B)                  ; \
    ASSIGN(B, T)                  ;
 
-#define SBOX6(A, B, C, D, T)       \
+#define SBOX_E6(A, B, C, D, T)     \
    XOR(A, B)                     ; \
    XOR(B, D)                     ; \
    NOT(D)                        ; \
@@ -151,7 +151,7 @@ START_LISTING(serp_asm.S)
    ASSIGN(B, D)                  ; \
    ASSIGN(D, T)                  ; \
 
-#define SBOX7(A, B, C, D, T)       \
+#define SBOX_E7(A, B, C, D, T)     \
    NOT(C)                        ; \
    ASSIGN(T, D)                  ; \
    AND(D, A)                     ; \
@@ -173,7 +173,7 @@ START_LISTING(serp_asm.S)
    ASSIGN(D, C)                  ; \
    ASSIGN(C, T)                  ;
 
-#define SBOX8(A, B, C, D, T)       \
+#define SBOX_E8(A, B, C, D, T)     \
    ASSIGN(T, B)                  ; \
    OR(B, C)                      ; \
    XOR(B, D)                     ; \
@@ -197,7 +197,201 @@ START_LISTING(serp_asm.S)
    ASSIGN(C, B)                  ; \
    ASSIGN(B, D)                  ; \
    ASSIGN(D, A)                  ; \
+   ASSIGN(A, T)                  ;
+
+#define SBOX_D1(A, B, C, D, T)     \
+   NOT(C)                        ; \
+   ASSIGN(T, B)                  ; \
+   OR(B, A)                      ; \
+   NOT(T)                        ; \
+   XOR(B, C)                     ; \
+   OR(C, T)                      ; \
+   XOR(B, D)                     ; \
+   XOR(A, T)                     ; \
+   XOR(C, A)                     ; \
+   AND(A, D)                     ; \
+   XOR(T, A)                     ; \
+   OR(A, B)                      ; \
+   XOR(A, C)                     ; \
+   XOR(D, T)                     ; \
+   XOR(C, B)                     ; \
+   XOR(D, A)                     ; \
+   XOR(D, B)                     ; \
+   AND(C, D)                     ; \
+   XOR(T, C)                     ; \
+   ASSIGN(C, B)                  ; \
+   ASSIGN(B, T)                  ;
+
+#define SBOX_D2(A, B, C, D, T)     \
+   ASSIGN(T, B)                  ; \
+   XOR(B, D)                     ; \
+   AND(D, B)                     ; \
+   XOR(T, C)                     ; \
+   XOR(D, A)                     ; \
+   OR(A, B)                      ; \
+   XOR(C, D)                     ; \
+   XOR(A, T)                     ; \
+   OR(A, C)                      ; \
+   XOR(B, D)                     ; \
+   XOR(A, B)                     ; \
+   OR(B, D)                      ; \
+   XOR(B, A)                     ; \
+   NOT(T)                        ; \
+   XOR(T, B)                     ; \
+   OR(B, A)                      ; \
+   XOR(B, A)                     ; \
+   OR(B, T)                      ; \
+   XOR(D, B)                     ; \
+   ASSIGN(B, A)                  ; \
    ASSIGN(A, T)                  ; \
+   ASSIGN(T, D)                  ; \
+   ASSIGN(D, C)                  ; \
+   ASSIGN(C, T)                  ;
+
+#define SBOX_D3(A, B, C, D, T)     \
+   XOR(C, D)                     ; \
+   XOR(D, A)                     ; \
+   ASSIGN(T, D)                  ; \
+   AND(D, C)                     ; \
+   XOR(D, B)                     ; \
+   OR(B, C)                      ; \
+   XOR(B, T)                     ; \
+   AND(T, D)                     ; \
+   XOR(C, D)                     ; \
+   AND(T, A)                     ; \
+   XOR(T, C)                     ; \
+   AND(C, B)                     ; \
+   OR(C, A)                      ; \
+   NOT(D)                        ; \
+   XOR(C, D)                     ; \
+   XOR(A, D)                     ; \
+   AND(A, B)                     ; \
+   XOR(D, T)                     ; \
+   XOR(D, A)                     ; \
+   ASSIGN(A, B)                  ; \
+   ASSIGN(B, T)                  ;
+
+   //busted
+#define SBOX_D4(A, B, C, D, T)     \
+   ASSIGN(T, C)                  ; \
+   XOR(C, B)                     ; \
+   XOR(A, C)                     ; \
+   AND(T, C)                     ; \
+   XOR(T, A)                     ; \
+   AND(A, B)                     ; \
+   XOR(B, D)                     ; \
+   OR(D, T)                      ; \
+   XOR(C, D)                     ; \
+   XOR(A, D)                     ; \
+   XOR(B, T)                     ; \
+   AND(D, C)                     ; \
+   XOR(D, B)                     ; \
+   XOR(B, A)                     ; \
+   OR(B, C)                      ; \
+   XOR(A, D)                     ; \
+   XOR(B, T)                     ; \
+   XOR(A, B)                     ; \
+   ASSIGN(T, A)                  ; \
+   ASSIGN(A, C)                  ; \
+   ASSIGN(C, D)                  ; \
+   ASSIGN(D, T)                  ;
+
+#define SBOX_D5(A, B, C, D, T)     \
+   ASSIGN(T, C)                  ; \
+   AND(C, D)                     ; \
+   XOR(C, B)                     ; \
+   OR(B, D)                      ; \
+   AND(B, A)                     ; \
+   XOR(T, C)                     ; \
+   XOR(T, B)                     ; \
+   AND(B, C)                     ; \
+   NOT(A)                        ; \
+   XOR(D, T)                     ; \
+   XOR(B, D)                     ; \
+   AND(D, A)                     ; \
+   XOR(D, C)                     ; \
+   XOR(A, B)                     ; \
+   AND(C, A)                     ; \
+   XOR(D, A)                     ; \
+   XOR(C, T)                     ; \
+   OR(C, D)                      ; \
+   XOR(D, A)                     ; \
+   XOR(C, B)                     ; \
+   ASSIGN(B, D)                  ; \
+   ASSIGN(D, T)                  ; \
+
+#define SBOX_D6(A, B, C, D, T)     \
+   NOT(B)                        ; \
+   ASSIGN(T, D)                  ; \
+   XOR(C, B)                     ; \
+   OR(D, A)                      ; \
+   XOR(D, C)                     ; \
+   OR(C, B)                      ; \
+   AND(C, A)                     ; \
+   XOR(T, D)                     ; \
+   XOR(C, T)                     ; \
+   OR(T, A)                      ; \
+   XOR(T, B)                     ; \
+   AND(B, C)                     ; \
+   XOR(B, D)                     ; \
+   XOR(T, C)                     ; \
+   AND(D, T)                     ; \
+   XOR(T, B)                     ; \
+   XOR(D, T)                     ; \
+   NOT(T)                        ; \
+   XOR(D, A)                     ; \
+   ASSIGN(A, B)                  ; \
+   ASSIGN(B, T)                  ; \
+   ASSIGN(T, D)                  ; \
+   ASSIGN(D, C)                  ; \
+   ASSIGN(C, T)                  ; \
+
+#define SBOX_D7(A, B, C, D, T)     \
+   XOR(A, C)                     ; \
+   ASSIGN(T, C)                  ; \
+   AND(C, A)                     ; \
+   XOR(T, D)                     ; \
+   NOT(C)                        ; \
+   XOR(D, B)                     ; \
+   XOR(C, D)                     ; \
+   OR(T, A)                      ; \
+   XOR(A, C)                     ; \
+   XOR(D, T)                     ; \
+   XOR(T, B)                     ; \
+   AND(B, D)                     ; \
+   XOR(B, A)                     ; \
+   XOR(A, D)                     ; \
+   OR(A, C)                      ; \
+   XOR(D, B)                     ; \
+   XOR(T, A)                     ; \
+   ASSIGN(A, B)                  ; \
+   ASSIGN(B, C)                  ; \
+   ASSIGN(C, T)                  ; \
+
+#define SBOX_D8(A, B, C, D, T)     \
+   ASSIGN(T, C)                  ; \
+   XOR(C, A)                     ; \
+   AND(A, D)                     ; \
+   OR(T, D)                      ; \
+   NOT(C)                        ; \
+   XOR(D, B)                     ; \
+   OR(B, A)                      ; \
+   XOR(A, C)                     ; \
+   AND(C, T)                     ; \
+   AND(D, T)                     ; \
+   XOR(B, C)                     ; \
+   XOR(C, A)                     ; \
+   OR(A, C)                      ; \
+   XOR(T, B)                     ; \
+   XOR(A, D)                     ; \
+   XOR(D, T)                     ; \
+   OR(T, A)                      ; \
+   XOR(D, C)                     ; \
+   XOR(T, C)                     ; \
+   ASSIGN(C, B)                  ; \
+   ASSIGN(B, A)                  ; \
+   ASSIGN(A, D)                  ; \
+   ASSIGN(D, T)                  ; \
 
 #define TRANSFORM(A, B, C, D, T)   \
    ROTL_IMM(A, 13)               ; \
@@ -219,18 +413,62 @@ START_LISTING(serp_asm.S)
    ROTL_IMM(A, 5)                ; \
    ROTL_IMM(C, 22)               ;
 
+#define I_TRANSFORM(A, B, C, D, T) \
+   ROTR_IMM(C, 22)               ; \
+   ROTR_IMM(A, 5)                ; \
+   ASSIGN(T, B)                  ; \
+   SHL_IMM(T, 7)                 ; \
+   XOR(C, D)                     ; \
+   XOR(C, T)                     ; \
+   XOR(A, B)                     ; \
+   XOR(A, D)                     ; \
+   ROTR_IMM(D, 7)                ; \
+   ROTR_IMM(B, 1)                ; \
+   ASSIGN(T, A)                  ; \
+   SHL_IMM(T, 3)                 ; \
+   XOR(D, T)                     ; \
+   XOR(D, C)                     ; \
+   XOR(B, C)                     ; \
+   XOR(B, A)                     ; \
+   ROTR_IMM(C, 3)                ; \
+   ROTR_IMM(A, 13)               ; \
+
 #define KEY_XOR(A, B, C, D, N)     \
    XOR(A, ARRAY4(EDI, (4*N  )))  ; \
    XOR(B, ARRAY4(EDI, (4*N+1)))  ; \
    XOR(C, ARRAY4(EDI, (4*N+2)))  ; \
    XOR(D, ARRAY4(EDI, (4*N+3)))  ; \
 
-START_FUNCTION(serpent_encrypt)
+START_FUNCTION(serp_sbox)
    SPILL_REGS()
 
 #define PUSHED 4
 
    ASSIGN(EBP, ARG(1)) /* input block */
+
+   ASSIGN(EAX, ARRAY4(EBP, 0))
+   ASSIGN(EBX, ARRAY4(EBP, 1))
+   ASSIGN(ECX, ARRAY4(EBP, 2))
+   ASSIGN(EDX, ARRAY4(EBP, 3))
+
+   SBOX_D4(EAX, EBX, ECX, EDX, EBP)
+
+   ASSIGN(EBP, ARG(1)) /* output block */
+   ASSIGN(ARRAY4(EBP, 0), EAX)
+   ASSIGN(ARRAY4(EBP, 1), EBX)
+   ASSIGN(ARRAY4(EBP, 2), ECX)
+   ASSIGN(ARRAY4(EBP, 3), EDX)
+
+   RESTORE_REGS()
+END_FUNCTION(serp_sbox)
+
+
+START_FUNCTION(serpent_encrypt)
+   SPILL_REGS()
+
+//#define PUSHED 4
+
+   ASSIGN(EBP, ARG(1)) /* input block */
    ASSIGN(EDI, ARG(3)) /* round keys */
 
    ASSIGN(EAX, ARRAY4(EBP, 0))
@@ -240,48 +478,48 @@ START_FUNCTION(serpent_encrypt)
 
    ZEROIZE(EBP)
 
-#define ROUND(A, B, C, D, T, N, SBOX) \
+#define E_ROUND(A, B, C, D, T, N, SBOX) \
    KEY_XOR(A, B, C, D, N) \
    SBOX(A, B, C, D, T) \
    TRANSFORM(A, B, C, D, T)
 
-   ROUND(EAX, EBX, ECX, EDX, EBP,  0, SBOX1)
-   ROUND(EAX, EBX, ECX, EDX, EBP,  1, SBOX2)
-   ROUND(EAX, EBX, ECX, EDX, EBP,  2, SBOX3)
-   ROUND(EAX, EBX, ECX, EDX, EBP,  3, SBOX4)
-   ROUND(EAX, EBX, ECX, EDX, EBP,  4, SBOX5)
-   ROUND(EAX, EBX, ECX, EDX, EBP,  5, SBOX6)
-   ROUND(EAX, EBX, ECX, EDX, EBP,  6, SBOX7)
-   ROUND(EAX, EBX, ECX, EDX, EBP,  7, SBOX8)
-
-   ROUND(EAX, EBX, ECX, EDX, EBP,  8, SBOX1)
-   ROUND(EAX, EBX, ECX, EDX, EBP,  9, SBOX2)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 10, SBOX3)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 11, SBOX4)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 12, SBOX5)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 13, SBOX6)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 14, SBOX7)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 15, SBOX8)
-
-   ROUND(EAX, EBX, ECX, EDX, EBP, 16, SBOX1)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 17, SBOX2)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 18, SBOX3)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 19, SBOX4)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 20, SBOX5)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 21, SBOX6)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 22, SBOX7)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 23, SBOX8)
-
-   ROUND(EAX, EBX, ECX, EDX, EBP, 24, SBOX1)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 25, SBOX2)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 26, SBOX3)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 27, SBOX4)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 28, SBOX5)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 29, SBOX6)
-   ROUND(EAX, EBX, ECX, EDX, EBP, 30, SBOX7)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP,  0, SBOX_E1)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP,  1, SBOX_E2)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP,  2, SBOX_E3)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP,  3, SBOX_E4)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP,  4, SBOX_E5)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP,  5, SBOX_E6)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP,  6, SBOX_E7)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP,  7, SBOX_E8)
+
+   E_ROUND(EAX, EBX, ECX, EDX, EBP,  8, SBOX_E1)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP,  9, SBOX_E2)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 10, SBOX_E3)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 11, SBOX_E4)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 12, SBOX_E5)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 13, SBOX_E6)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 14, SBOX_E7)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 15, SBOX_E8)
+
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 16, SBOX_E1)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 17, SBOX_E2)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 18, SBOX_E3)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 19, SBOX_E4)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 20, SBOX_E5)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 21, SBOX_E6)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 22, SBOX_E7)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 23, SBOX_E8)
+
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 24, SBOX_E1)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 25, SBOX_E2)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 26, SBOX_E3)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 27, SBOX_E4)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 28, SBOX_E5)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 29, SBOX_E6)
+   E_ROUND(EAX, EBX, ECX, EDX, EBP, 30, SBOX_E7)
 
    KEY_XOR(EAX, EBX, ECX, EDX, 31)
-   SBOX8(EAX, EBX, ECX, EDX, EBP)
+   SBOX_E8(EAX, EBX, ECX, EDX, EBP)
    KEY_XOR(EAX, EBX, ECX, EDX, 32)
 
    ASSIGN(EBP, ARG(2)) /* output block */
@@ -292,3 +530,70 @@ START_FUNCTION(serpent_encrypt)
 
    RESTORE_REGS()
 END_FUNCTION(serpent_encrypt)
+
+START_FUNCTION(serpent_decrypt)
+   SPILL_REGS()
+
+   ASSIGN(EBP, ARG(1)) /* input block */
+   ASSIGN(EDI, ARG(3)) /* round keys */
+
+   ASSIGN(EAX, ARRAY4(EBP, 0))
+   ASSIGN(EBX, ARRAY4(EBP, 1))
+   ASSIGN(ECX, ARRAY4(EBP, 2))
+   ASSIGN(EDX, ARRAY4(EBP, 3))
+
+   ZEROIZE(EBP)
+
+#define D_ROUND(A, B, C, D, T, N, SBOX) \
+   I_TRANSFORM(A, B, C, D, T) \
+   SBOX(A, B, C, D, T) \
+   KEY_XOR(A, B, C, D, N) \
+
+   KEY_XOR(EAX, EBX, ECX, EDX, 32)
+   SBOX_D8(EAX, EBX, ECX, EDX, EBP)
+   KEY_XOR(EAX, EBX, ECX, EDX, 31)
+
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 30, SBOX_D7)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 29, SBOX_D6)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 28, SBOX_D5)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 27, SBOX_D4)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 26, SBOX_D3)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 25, SBOX_D2)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 24, SBOX_D1)
+
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 23, SBOX_D8)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 22, SBOX_D7)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 21, SBOX_D6)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 20, SBOX_D5)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 19, SBOX_D4)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 18, SBOX_D3)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 17, SBOX_D2)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 16, SBOX_D1)
+
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 15, SBOX_D8)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 14, SBOX_D7)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 13, SBOX_D6)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 12, SBOX_D5)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 11, SBOX_D4)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP, 10, SBOX_D3)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP,  9, SBOX_D2)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP,  8, SBOX_D1)
+
+   D_ROUND(EAX, EBX, ECX, EDX, EBP,  7, SBOX_D8)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP,  6, SBOX_D7)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP,  5, SBOX_D6)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP,  4, SBOX_D5)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP,  3, SBOX_D4)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP,  2, SBOX_D3)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP,  1, SBOX_D2)
+   D_ROUND(EAX, EBX, ECX, EDX, EBP,  0, SBOX_D1)
+
+   ASSIGN(EBP, ARG(2)) /* output block */
+   ASSIGN(ARRAY4(EBP, 0), EAX)
+   ASSIGN(ARRAY4(EBP, 1), EBX)
+   ASSIGN(ARRAY4(EBP, 2), ECX)
+   ASSIGN(ARRAY4(EBP, 3), EDX)
+
+   RESTORE_REGS()
+END_FUNCTION(serpent_decrypt)
+
diff --git a/modules/alg_ia32/serpent.cpp b/modules/alg_ia32/serpent.cpp
index aaa666ea7..c6db285c1 100644
--- a/modules/alg_ia32/serpent.cpp
+++ b/modules/alg_ia32/serpent.cpp
@@ -105,127 +105,10 @@ inline void SBoxE8(u32bit& B0, u32bit& B1, u32bit& B2, u32bit& B3)
    B0 = T4; B1 = T3; B2 = T1; B3 = T0;
    }
 
-/*************************************************
-* Serpent Decryption S-Box 1                     *
-*************************************************/
-inline void SBoxD1(u32bit& B0, u32bit& B1, u32bit& B2, u32bit& B3)
-   {
-   u32bit T0 = B0, T1 = B1, T2 = B2, T3 = B3, T4;
-   T2 = ~T2; T4  = T1; T1 |= T0; T4 = ~T4; T1 ^= T2; T2 |= T4; T1 ^= T3;
-   T0 ^= T4; T2 ^= T0; T0 &= T3; T4 ^= T0; T0 |= T1; T0 ^= T2; T3 ^= T4;
-   T2 ^= T1; T3 ^= T0; T3 ^= T1; T2 &= T3; T4 ^= T2;
-   B0 = T0; B1 = T4; B2 = T1; B3 = T3;
-   }
-
-/*************************************************
-* Serpent Decryption S-Box 2                     *
-*************************************************/
-inline void SBoxD2(u32bit& B0, u32bit& B1, u32bit& B2, u32bit& B3)
-   {
-   u32bit T0 = B0, T1 = B1, T2 = B2, T3 = B3, T4;
-   T4  = T1; T1 ^= T3; T3 &= T1; T4 ^= T2; T3 ^= T0; T0 |= T1; T2 ^= T3;
-   T0 ^= T4; T0 |= T2; T1 ^= T3; T0 ^= T1; T1 |= T3; T1 ^= T0; T4 = ~T4;
-   T4 ^= T1; T1 |= T0; T1 ^= T0; T1 |= T4; T3 ^= T1;
-   B0 = T4; B1 = T0; B2 = T3; B3 = T2;
-   }
-
-/*************************************************
-* Serpent Decryption S-Box 3                     *
-*************************************************/
-inline void SBoxD3(u32bit& B0, u32bit& B1, u32bit& B2, u32bit& B3)
-   {
-   u32bit T0 = B0, T1 = B1, T2 = B2, T3 = B3, T4;
-   T2 ^= T3; T3 ^= T0; T4  = T3; T3 &= T2; T3 ^= T1; T1 |= T2; T1 ^= T4;
-   T4 &= T3; T2 ^= T3; T4 &= T0; T4 ^= T2; T2 &= T1; T2 |= T0; T3 = ~T3;
-   T2 ^= T3; T0 ^= T3; T0 &= T1; T3 ^= T4; T3 ^= T0;
-   B0 = T1; B1 = T4; B2 = T2; B3 = T3;
-   }
-
-/*************************************************
-* Serpent Decryption S-Box 4                     *
-*************************************************/
-inline void SBoxD4(u32bit& B0, u32bit& B1, u32bit& B2, u32bit& B3)
-   {
-   u32bit T0 = B0, T1 = B1, T2 = B2, T3 = B3, T4;
-   T4  = T2; T2 ^= T1; T0 ^= T2; T4 &= T2; T4 ^= T0; T0 &= T1; T1 ^= T3;
-   T3 |= T4; T2 ^= T3; T0 ^= T3; T1 ^= T4; T3 &= T2; T3 ^= T1; T1 ^= T0;
-   T1 |= T2; T0 ^= T3; T1 ^= T4; T0 ^= T1;
-   B0 = T2; B1 = T1; B2 = T3; B3 = T0;
-   }
-
-/*************************************************
-* Serpent Decryption S-Box 5                     *
-*************************************************/
-inline void SBoxD5(u32bit& B0, u32bit& B1, u32bit& B2, u32bit& B3)
-   {
-   u32bit T0 = B0, T1 = B1, T2 = B2, T3 = B3, T4;
-   T4  = T2; T2 &= T3; T2 ^= T1; T1 |= T3; T1 &= T0; T4 ^= T2; T4 ^= T1;
-   T1 &= T2; T0 = ~T0; T3 ^= T4; T1 ^= T3; T3 &= T0; T3 ^= T2; T0 ^= T1;
-   T2 &= T0; T3 ^= T0; T2 ^= T4; T2 |= T3; T3 ^= T0; T2 ^= T1;
-   B0 = T0; B1 = T3; B2 = T2; B3 = T4;
-   }
-
-/*************************************************
-* Serpent Decryption S-Box 6                     *
-*************************************************/
-inline void SBoxD6(u32bit& B0, u32bit& B1, u32bit& B2, u32bit& B3)
-   {
-   u32bit T0 = B0, T1 = B1, T2 = B2, T3 = B3, T4;
-   T1 = ~T1; T4  = T3; T2 ^= T1; T3 |= T0; T3 ^= T2; T2 |= T1; T2 &= T0;
-   T4 ^= T3; T2 ^= T4; T4 |= T0; T4 ^= T1; T1 &= T2; T1 ^= T3; T4 ^= T2;
-   T3 &= T4; T4 ^= T1; T3 ^= T4; T4 = ~T4; T3 ^= T0;
-   B0 = T1; B1 = T4; B2 = T3; B3 = T2;
-   }
-
-/*************************************************
-* Serpent Decryption S-Box 7                     *
-*************************************************/
-inline void SBoxD7(u32bit& B0, u32bit& B1, u32bit& B2, u32bit& B3)
-   {
-   u32bit T0 = B0, T1 = B1, T2 = B2, T3 = B3, T4;
-   T0 ^= T2; T4  = T2; T2 &= T0; T4 ^= T3; T2 = ~T2; T3 ^= T1; T2 ^= T3;
-   T4 |= T0; T0 ^= T2; T3 ^= T4; T4 ^= T1; T1 &= T3; T1 ^= T0; T0 ^= T3;
-   T0 |= T2; T3 ^= T1; T4 ^= T0;
-   B0 = T1; B1 = T2; B2 = T4; B3 = T3;
-   }
-
-/*************************************************
-* Serpent Decryption S-Box 8                     *
-*************************************************/
-inline void SBoxD8(u32bit& B0, u32bit& B1, u32bit& B2, u32bit& B3)
-   {
-   u32bit T0 = B0, T1 = B1, T2 = B2, T3 = B3, T4;
-   T4  = T2; T2 ^= T0; T0 &= T3; T4 |= T3; T2 = ~T2; T3 ^= T1; T1 |= T0;
-   T0 ^= T2; T2 &= T4; T3 &= T4; T1 ^= T2; T2 ^= T0; T0 |= T2; T4 ^= T1;
-   T0 ^= T3; T3 ^= T4; T4 |= T0; T3 ^= T2; T4 ^= T2;
-   B0 = T3; B1 = T0; B2 = T1; B3 = T4;
-   }
-
-/*************************************************
-* Serpent's Inverse Linear Transformation        *
-*************************************************/
-inline void i_transform(u32bit& B0, u32bit& B1, u32bit& B2, u32bit& B3)
-   {
-   B2  = rotate_right(B2, 22);  B0  = rotate_right(B0, 5);
-   B2 ^= B3 ^ (B1 << 7);        B0 ^= B1 ^ B3;
-   B3  = rotate_right(B3, 7);   B1  = rotate_right(B1, 1);
-   B3 ^= B2 ^ (B0 << 3);        B1 ^= B0 ^ B2;
-   B2  = rotate_right(B2, 3);   B0  = rotate_right(B0, 13);
-   }
-
 }
 
-/*************************************************
-* XOR a key block with a data block              *
-*************************************************/
-inline void Serpent::key_xor(u32bit round, u32bit& B0, u32bit& B1,
-                                           u32bit& B2, u32bit& B3) const
-   {
-   B0 ^= round_key[4*round  ]; B1 ^= round_key[4*round+1];
-   B2 ^= round_key[4*round+2]; B3 ^= round_key[4*round+3];
-   }
-
 extern "C" void serpent_encrypt(const byte[16], byte[16], const u32bit[132]);
+extern "C" void serpent_decrypt(const byte[16], byte[16], const u32bit[132]);
 
 /*************************************************
 * Serpent Encryption                             *
@@ -240,50 +123,7 @@ void Serpent::enc(const byte in[], byte out[]) const
 *************************************************/
 void Serpent::dec(const byte in[], byte out[]) const
    {
-   u32bit B0 = make_u32bit(in[ 3], in[ 2], in[ 1], in[ 0]),
-          B1 = make_u32bit(in[ 7], in[ 6], in[ 5], in[ 4]),
-          B2 = make_u32bit(in[11], in[10], in[ 9], in[ 8]),
-          B3 = make_u32bit(in[15], in[14], in[13], in[12]);
-   key_xor(32,B0,B1,B2,B3);  SBoxD8(B0,B1,B2,B3); key_xor(31,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD7(B0,B1,B2,B3); key_xor(30,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD6(B0,B1,B2,B3); key_xor(29,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD5(B0,B1,B2,B3); key_xor(28,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD4(B0,B1,B2,B3); key_xor(27,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD3(B0,B1,B2,B3); key_xor(26,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD2(B0,B1,B2,B3); key_xor(25,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD1(B0,B1,B2,B3); key_xor(24,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD8(B0,B1,B2,B3); key_xor(23,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD7(B0,B1,B2,B3); key_xor(22,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD6(B0,B1,B2,B3); key_xor(21,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD5(B0,B1,B2,B3); key_xor(20,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD4(B0,B1,B2,B3); key_xor(19,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD3(B0,B1,B2,B3); key_xor(18,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD2(B0,B1,B2,B3); key_xor(17,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD1(B0,B1,B2,B3); key_xor(16,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD8(B0,B1,B2,B3); key_xor(15,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD7(B0,B1,B2,B3); key_xor(14,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD6(B0,B1,B2,B3); key_xor(13,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD5(B0,B1,B2,B3); key_xor(12,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD4(B0,B1,B2,B3); key_xor(11,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD3(B0,B1,B2,B3); key_xor(10,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD2(B0,B1,B2,B3); key_xor( 9,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD1(B0,B1,B2,B3); key_xor( 8,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD8(B0,B1,B2,B3); key_xor( 7,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD7(B0,B1,B2,B3); key_xor( 6,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD6(B0,B1,B2,B3); key_xor( 5,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD5(B0,B1,B2,B3); key_xor( 4,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD4(B0,B1,B2,B3); key_xor( 3,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD3(B0,B1,B2,B3); key_xor( 2,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD2(B0,B1,B2,B3); key_xor( 1,B0,B1,B2,B3);
-   i_transform(B0,B1,B2,B3); SBoxD1(B0,B1,B2,B3); key_xor( 0,B0,B1,B2,B3);
-   out[ 0] = get_byte(3, B0); out[ 1] = get_byte(2, B0);
-   out[ 2] = get_byte(1, B0); out[ 3] = get_byte(0, B0);
-   out[ 4] = get_byte(3, B1); out[ 5] = get_byte(2, B1);
-   out[ 6] = get_byte(1, B1); out[ 7] = get_byte(0, B1);
-   out[ 8] = get_byte(3, B2); out[ 9] = get_byte(2, B2);
-   out[10] = get_byte(1, B2); out[11] = get_byte(0, B2);
-   out[12] = get_byte(3, B3); out[13] = get_byte(2, B3);
-   out[14] = get_byte(1, B3); out[15] = get_byte(0, B3);
+   serpent_decrypt(in, out, round_key);
    }
 
 /*************************************************
diff --git a/modules/alg_ia32/serpent.h b/modules/alg_ia32/serpent.h
new file mode 100644
index 000000000..85e0345d0
--- /dev/null
+++ b/modules/alg_ia32/serpent.h
@@ -0,0 +1,33 @@
+/*************************************************
+* Serpent Header File                            *
+* (C) 1999-2006 The Botan Project                *
+*************************************************/
+
+#ifndef BOTAN_SERPENT_H__
+#define BOTAN_SERPENT_H__
+
+#include <botan/base.h>
+
+namespace Botan {
+
+/*************************************************
+* Serpent                                        *
+*************************************************/
+class Serpent : public BlockCipher
+   {
+   public:
+      void clear() throw() { round_key.clear(); }
+      std::string name() const { return "Serpent"; }
+      BlockCipher* clone() const { return new Serpent; }
+      Serpent() : BlockCipher(16, 16, 32, 8) {}
+   private:
+      void enc(const byte[], byte[]) const;
+      void dec(const byte[], byte[]) const;
+      void key(const byte[], u32bit);
+
+      SecureBuffer<u32bit, 132> round_key;
+   };
+
+}
+
+#endif