Have the expansion loop in the key schedule take advantage of free

registers to load words we will need in advance.

2 files changed, 17 insertions, 12 deletions

diff --git a/modules/alg_ia32/serp_asm.S b/modules/alg_ia32/serp_asm.S
index e6c8c3cb3..30c154516 100644
--- a/modules/alg_ia32/serp_asm.S
+++ b/modules/alg_ia32/serp_asm.S
@@ -443,7 +443,6 @@ START_LISTING(serp_asm.S)
 *************************************************/
 START_FUNCTION(serpent_encrypt)
    SPILL_REGS()
-
 #define PUSHED 4
 
    ASSIGN(EBP, ARG(1)) /* input block */
@@ -507,6 +506,7 @@ START_FUNCTION(serpent_encrypt)
    ASSIGN(ARRAY4(EBP, 3), EDX)
 
    RESTORE_REGS()
+#undef PUSHED
 END_FUNCTION(serpent_encrypt)
 
 /*************************************************
@@ -514,6 +514,7 @@ END_FUNCTION(serpent_encrypt)
 *************************************************/
 START_FUNCTION(serpent_decrypt)
    SPILL_REGS()
+#define PUSHED 4
 
    ASSIGN(EBP, ARG(1)) /* input block */
    ASSIGN(EDI, ARG(3)) /* round keys */
@@ -576,6 +577,7 @@ START_FUNCTION(serpent_decrypt)
    ASSIGN(ARRAY4(EBP, 3), EDX)
 
    RESTORE_REGS()
+#undef PUSHED
 END_FUNCTION(serpent_decrypt)
 
 /*************************************************
@@ -583,21 +585,25 @@ END_FUNCTION(serpent_decrypt)
 *************************************************/
 START_FUNCTION(serpent_key_schedule)
    SPILL_REGS()
+#define PUSHED 4
 
    ASSIGN(EDI, ARG(1)) /* round keys */
    ASSIGN(ESI, IMM(8))
    ADD_IMM(EDI, 32)
 
 START_LOOP(.EXPANSION)
-
-   ASSIGN(EAX, ESI)
-   SUB_IMM(EAX, 8)
-
-   XOR(EAX, ARRAY4(EDI, -1))
-   XOR(EAX, ARRAY4(EDI, -3))
-   XOR(EAX, ARRAY4(EDI, -5))
-   XOR(EAX, ARRAY4(EDI, -8))
-   XOR(EAX, IMM(0x9E3779B9))
+   ASSIGN(EAX, ARRAY4(EDI, -1))
+   ASSIGN(EBX, ARRAY4(EDI, -3))
+   ASSIGN(ECX, ARRAY4(EDI, -5))
+   ASSIGN(EDX, ARRAY4(EDI, -8))
+
+   ASSIGN(EBP, ESI)
+   SUB_IMM(EBP, 8)
+   XOR(EBP, IMM(0x9E3779B9))
+   XOR(EAX, EBX)
+   XOR(ECX, EDX)
+   XOR(EAX, EBP)
+   XOR(EAX, ECX)
 
    ROTL_IMM(EAX, 11)
 
@@ -607,7 +613,6 @@ START_LOOP(.EXPANSION)
    ADD_IMM(ESI, 1)
 LOOP_UNTIL(ESI, IMM(140), .EXPANSION)
 
-
    ASSIGN(EDI, ARG(1)) /* round keys */
 
 #define LOAD_AND_SBOX(MSG, SBOX)          \
@@ -660,5 +665,6 @@ LOOP_UNTIL(ESI, IMM(140), .EXPANSION)
    LOAD_AND_SBOX(32, SBOX_E4)
 
    RESTORE_REGS()
+#undef PUSHED
 END_FUNCTION(serpent_key_schedule)
 
diff --git a/modules/alg_ia32/serpent.cpp b/modules/alg_ia32/serpent.cpp
index 964b677a1..fd489574b 100644
--- a/modules/alg_ia32/serpent.cpp
+++ b/modules/alg_ia32/serpent.cpp
@@ -43,7 +43,6 @@ void Serpent::key(const byte key[], u32bit length)
    W[length / 4] |= u32bit(1) << ((length%4)*8);
 
    serpent_key_schedule(W);
-
    round_key.copy(W + 8, 132);
    }