Cleanups, and move the initial memory access to the beginning of each

MD5 round in an attempt to hide the latency a bit

2 files changed, 77 insertions, 52 deletions

diff --git a/modules/alg_ia32/md5core.S b/modules/alg_ia32/md5core.S
index a801a3681..7bc7ec99d 100644
--- a/modules/alg_ia32/md5core.S
+++ b/modules/alg_ia32/md5core.S
@@ -32,94 +32,119 @@ START_LOOP(.LOAD_INPUT)
    ASSIGN(ARRAY4_INDIRECT(EDI,ESI,-2), ECX)
    ASSIGN(ARRAY4_INDIRECT(EDI,ESI,-1), EDX)
 LOOP_UNTIL(ESI, IMM(16), .LOAD_INPUT)
-   
+
    ASSIGN(EBP, ARG(1))
    ASSIGN(EAX, ARRAY4(EBP, 0))
    ASSIGN(EBX, ARRAY4(EBP, 1))
    ASSIGN(ECX, ARRAY4(EBP, 2))
    ASSIGN(EDX, ARRAY4(EBP, 3))
 
-#define T EBP
-   
-#define FF(A, B, C, D, MSG, S, MAGIC)    \
+#define FF(A, B, C, D, T, MSG, S, MAGIC) \
+   ADD(A, ARRAY4(EDI, MSG))            ; \
    ASSIGN(T, C)                        ; \
    XOR(T, D)                           ; \
    AND(T, B)                           ; \
    XOR(T, D)                           ; \
-   ADD(A, ARRAY4(EDI, MSG))            ; \
    ADD3_IMM(A, T, MAGIC)               ; \
    ROTL_IMM(A, S)                      ; \
    ADD(A, B)                           ;
 
-#define GG(A, B, C, D, MSG, S, MAGIC)    \
+#define GG(A, B, C, D, T, MSG, S, MAGIC) \
+   ADD(A, ARRAY4(EDI, MSG))            ; \
    ASSIGN(T, B)                        ; \
    XOR(T, C)                           ; \
    AND(T, D)                           ; \
    XOR(T, C)                           ; \
-   ADD(A, ARRAY4(EDI, MSG))            ; \
    ADD3_IMM(A, T, MAGIC)               ; \
    ROTL_IMM(A, S)                      ; \
    ADD(A, B)                           ;
 
-#define HH(A, B, C, D, MSG, S, MAGIC)    \
+#define HH(A, B, C, D, T, MSG, S, MAGIC) \
+   ADD(A, ARRAY4(EDI, MSG))            ; \
    ASSIGN(T, B)                        ; \
    XOR(T, C)                           ; \
    XOR(T, D)                           ; \
-   ADD(A, ARRAY4(EDI, MSG))            ; \
    ADD3_IMM(A, T, MAGIC)               ; \
    ROTL_IMM(A, S)                      ; \
    ADD(A, B)                           ;
 
-#define II(A, B, C, D, MSG, S, MAGIC)    \
+#define II(A, B, C, D, T, MSG, S, MAGIC) \
+   ADD(A, ARRAY4(EDI, MSG))            ; \
    ASSIGN(T, D)                        ; \
    NOT(T)                              ; \
    OR(T, B)                            ; \
    XOR(T, C)                           ; \
-   ADD(A, ARRAY4(EDI, MSG))            ; \
    ADD3_IMM(A, T, MAGIC)               ; \
    ROTL_IMM(A, S)                      ; \
    ADD(A, B)                           ;
 
-#define A EAX
-#define B EBX
-#define C ECX
-#define D EDX
-   
-   FF(A,B,C,D, 0, 7,0xD76AA478);   FF(D,A,B,C, 1,12,0xE8C7B756);
-   FF(C,D,A,B, 2,17,0x242070DB);   FF(B,C,D,A, 3,22,0xC1BDCEEE);
-   FF(A,B,C,D, 4, 7,0xF57C0FAF);   FF(D,A,B,C, 5,12,0x4787C62A);
-   FF(C,D,A,B, 6,17,0xA8304613);   FF(B,C,D,A, 7,22,0xFD469501);
-   FF(A,B,C,D, 8, 7,0x698098D8);   FF(D,A,B,C, 9,12,0x8B44F7AF);
-   FF(C,D,A,B,10,17,0xFFFF5BB1);   FF(B,C,D,A,11,22,0x895CD7BE);
-   FF(A,B,C,D,12, 7,0x6B901122);   FF(D,A,B,C,13,12,0xFD987193);
-   FF(C,D,A,B,14,17,0xA679438E);   FF(B,C,D,A,15,22,0x49B40821);
-
-   GG(A,B,C,D, 1, 5,0xF61E2562);   GG(D,A,B,C, 6, 9,0xC040B340);
-   GG(C,D,A,B,11,14,0x265E5A51);   GG(B,C,D,A, 0,20,0xE9B6C7AA);
-   GG(A,B,C,D, 5, 5,0xD62F105D);   GG(D,A,B,C,10, 9,0x02441453);
-   GG(C,D,A,B,15,14,0xD8A1E681);   GG(B,C,D,A, 4,20,0xE7D3FBC8);
-   GG(A,B,C,D, 9, 5,0x21E1CDE6);   GG(D,A,B,C,14, 9,0xC33707D6);
-   GG(C,D,A,B, 3,14,0xF4D50D87);   GG(B,C,D,A, 8,20,0x455A14ED);
-   GG(A,B,C,D,13, 5,0xA9E3E905);   GG(D,A,B,C, 2, 9,0xFCEFA3F8);
-   GG(C,D,A,B, 7,14,0x676F02D9);   GG(B,C,D,A,12,20,0x8D2A4C8A);
-
-   HH(A,B,C,D, 5, 4,0xFFFA3942);   HH(D,A,B,C, 8,11,0x8771F681);
-   HH(C,D,A,B,11,16,0x6D9D6122);   HH(B,C,D,A,14,23,0xFDE5380C);
-   HH(A,B,C,D, 1, 4,0xA4BEEA44);   HH(D,A,B,C, 4,11,0x4BDECFA9);
-   HH(C,D,A,B, 7,16,0xF6BB4B60);   HH(B,C,D,A,10,23,0xBEBFBC70);
-   HH(A,B,C,D,13, 4,0x289B7EC6);   HH(D,A,B,C, 0,11,0xEAA127FA);
-   HH(C,D,A,B, 3,16,0xD4EF3085);   HH(B,C,D,A, 6,23,0x04881D05);
-   HH(A,B,C,D, 9, 4,0xD9D4D039);   HH(D,A,B,C,12,11,0xE6DB99E5);
-   HH(C,D,A,B,15,16,0x1FA27CF8);   HH(B,C,D,A, 2,23,0xC4AC5665);
-
-   II(A,B,C,D, 0, 6,0xF4292244);   II(D,A,B,C, 7,10,0x432AFF97);
-   II(C,D,A,B,14,15,0xAB9423A7);   II(B,C,D,A, 5,21,0xFC93A039);
-   II(A,B,C,D,12, 6,0x655B59C3);   II(D,A,B,C, 3,10,0x8F0CCC92);
-   II(C,D,A,B,10,15,0xFFEFF47D);   II(B,C,D,A, 1,21,0x85845DD1);
-   II(A,B,C,D, 8, 6,0x6FA87E4F);   II(D,A,B,C,15,10,0xFE2CE6E0);
-   II(C,D,A,B, 6,15,0xA3014314);   II(B,C,D,A,13,21,0x4E0811A1);
-   II(A,B,C,D, 4, 6,0xF7537E82);   II(D,A,B,C,11,10,0xBD3AF235);
-   II(C,D,A,B, 2,15,0x2AD7D2BB);   II(B,C,D,A, 9,21,0xEB86D391);
+   FF(EAX,EBX,ECX,EDX,ESI, 0, 7,0xD76AA478);
+   FF(EDX,EAX,EBX,ECX,ESI, 1,12,0xE8C7B756);
+   FF(ECX,EDX,EAX,EBX,ESI, 2,17,0x242070DB);
+   FF(EBX,ECX,EDX,EAX,ESI, 3,22,0xC1BDCEEE);
+   FF(EAX,EBX,ECX,EDX,ESI, 4, 7,0xF57C0FAF);
+   FF(EDX,EAX,EBX,ECX,ESI, 5,12,0x4787C62A);
+   FF(ECX,EDX,EAX,EBX,ESI, 6,17,0xA8304613);
+   FF(EBX,ECX,EDX,EAX,ESI, 7,22,0xFD469501);
+   FF(EAX,EBX,ECX,EDX,ESI, 8, 7,0x698098D8);
+   FF(EDX,EAX,EBX,ECX,ESI, 9,12,0x8B44F7AF);
+   FF(ECX,EDX,EAX,EBX,ESI,10,17,0xFFFF5BB1);
+   FF(EBX,ECX,EDX,EAX,ESI,11,22,0x895CD7BE);
+   FF(EAX,EBX,ECX,EDX,ESI,12, 7,0x6B901122);
+   FF(EDX,EAX,EBX,ECX,ESI,13,12,0xFD987193);
+   FF(ECX,EDX,EAX,EBX,ESI,14,17,0xA679438E);
+   FF(EBX,ECX,EDX,EAX,ESI,15,22,0x49B40821);
+
+   GG(EAX,EBX,ECX,EDX,ESI, 1, 5,0xF61E2562);
+   GG(EDX,EAX,EBX,ECX,ESI, 6, 9,0xC040B340);
+   GG(ECX,EDX,EAX,EBX,ESI,11,14,0x265E5A51);
+   GG(EBX,ECX,EDX,EAX,ESI, 0,20,0xE9B6C7AA);
+   GG(EAX,EBX,ECX,EDX,ESI, 5, 5,0xD62F105D);
+   GG(EDX,EAX,EBX,ECX,ESI,10, 9,0x02441453);
+   GG(ECX,EDX,EAX,EBX,ESI,15,14,0xD8A1E681);
+   GG(EBX,ECX,EDX,EAX,ESI, 4,20,0xE7D3FBC8);
+   GG(EAX,EBX,ECX,EDX,ESI, 9, 5,0x21E1CDE6);
+   GG(EDX,EAX,EBX,ECX,ESI,14, 9,0xC33707D6);
+   GG(ECX,EDX,EAX,EBX,ESI, 3,14,0xF4D50D87);
+   GG(EBX,ECX,EDX,EAX,ESI, 8,20,0x455A14ED);
+   GG(EAX,EBX,ECX,EDX,ESI,13, 5,0xA9E3E905);
+   GG(EDX,EAX,EBX,ECX,ESI, 2, 9,0xFCEFA3F8);
+   GG(ECX,EDX,EAX,EBX,ESI, 7,14,0x676F02D9);
+   GG(EBX,ECX,EDX,EAX,ESI,12,20,0x8D2A4C8A);
+
+   HH(EAX,EBX,ECX,EDX,ESI, 5, 4,0xFFFA3942);
+   HH(EDX,EAX,EBX,ECX,ESI, 8,11,0x8771F681);
+   HH(ECX,EDX,EAX,EBX,ESI,11,16,0x6D9D6122);
+   HH(EBX,ECX,EDX,EAX,ESI,14,23,0xFDE5380C);
+   HH(EAX,EBX,ECX,EDX,ESI, 1, 4,0xA4BEEA44);
+   HH(EDX,EAX,EBX,ECX,ESI, 4,11,0x4BDECFA9);
+   HH(ECX,EDX,EAX,EBX,ESI, 7,16,0xF6BB4B60);
+   HH(EBX,ECX,EDX,EAX,ESI,10,23,0xBEBFBC70);
+   HH(EAX,EBX,ECX,EDX,ESI,13, 4,0x289B7EC6);
+   HH(EDX,EAX,EBX,ECX,ESI, 0,11,0xEAA127FA);
+   HH(ECX,EDX,EAX,EBX,ESI, 3,16,0xD4EF3085);
+   HH(EBX,ECX,EDX,EAX,ESI, 6,23,0x04881D05);
+   HH(EAX,EBX,ECX,EDX,ESI, 9, 4,0xD9D4D039);
+   HH(EDX,EAX,EBX,ECX,ESI,12,11,0xE6DB99E5);
+   HH(ECX,EDX,EAX,EBX,ESI,15,16,0x1FA27CF8);
+   HH(EBX,ECX,EDX,EAX,ESI, 2,23,0xC4AC5665);
+
+   II(EAX,EBX,ECX,EDX,ESI, 0, 6,0xF4292244);
+   II(EDX,EAX,EBX,ECX,ESI, 7,10,0x432AFF97);
+   II(ECX,EDX,EAX,EBX,ESI,14,15,0xAB9423A7);
+   II(EBX,ECX,EDX,EAX,ESI, 5,21,0xFC93A039);
+   II(EAX,EBX,ECX,EDX,ESI,12, 6,0x655B59C3);
+   II(EDX,EAX,EBX,ECX,ESI, 3,10,0x8F0CCC92);
+   II(ECX,EDX,EAX,EBX,ESI,10,15,0xFFEFF47D);
+   II(EBX,ECX,EDX,EAX,ESI, 1,21,0x85845DD1);
+   II(EAX,EBX,ECX,EDX,ESI, 8, 6,0x6FA87E4F);
+   II(EDX,EAX,EBX,ECX,ESI,15,10,0xFE2CE6E0);
+   II(ECX,EDX,EAX,EBX,ESI, 6,15,0xA3014314);
+   II(EBX,ECX,EDX,EAX,ESI,13,21,0x4E0811A1);
+   II(EAX,EBX,ECX,EDX,ESI, 4, 6,0xF7537E82);
+   II(EDX,EAX,EBX,ECX,ESI,11,10,0xBD3AF235);
+   II(ECX,EDX,EAX,EBX,ESI, 2,15,0x2AD7D2BB);
+   II(EBX,ECX,EDX,EAX,ESI, 9,21,0xEB86D391);
 
    ASSIGN(EBP, ARG(1))
    ADD(ARRAY4(EBP, 0), EAX)
diff --git a/modules/alg_ia32/sha1core.S b/modules/alg_ia32/sha1core.S
index 45f95947d..191686a7d 100644
--- a/modules/alg_ia32/sha1core.S
+++ b/modules/alg_ia32/sha1core.S
@@ -100,7 +100,7 @@ LOOP_UNTIL(ESI, IMM(80), .EXPANSION)
    ADD(E, A)                    ; \
    ADD3_IMM(E, T, MAGIC1)       ; \
    ROTR_IMM(A, 5)               ;
-   
+
 #define F2_OR_F4(A, B, C, D, E, T, MSG, MAGIC) \
    ROTL_IMM(A, 5)               ; \
    ADD(E, ARRAY4(EDI, MSG))     ; \