Split up asm modules into appropriate (topic-specific) modules, eg

hash/sha1_amd64 and cipher/serpent_ia32.

Remaining code in asm/ dir is for BigInt, so rename to bigint/ in prep for
all (or most) of BigInt being modularized.

2 files changed, 294 insertions, 0 deletions

diff --git a/src/hash/sha1_ia32/sha160.cpp b/src/hash/sha1_ia32/sha160.cpp
new file mode 100644
index 000000000..7725541d5
--- /dev/null
+++ b/src/hash/sha1_ia32/sha160.cpp
@@ -0,0 +1,52 @@
+/*************************************************
+* SHA-160 Source File                            *
+* (C) 1999-2007 Jack Lloyd                       *
+*************************************************/
+
+#include <botan/sha160.h>
+#include <botan/loadstor.h>
+
+namespace Botan {
+
+extern "C" void botan_sha160_asm_ia32(u32bit[5], const byte[64], u32bit[81]);
+
+/*************************************************
+* SHA-160 Compression Function                   *
+*************************************************/
+void SHA_160::hash(const byte input[])
+   {
+   botan_sha160_asm_ia32(digest, input, W);
+   }
+
+/*************************************************
+* Copy out the digest                            *
+*************************************************/
+void SHA_160::copy_out(byte output[])
+   {
+   for(u32bit j = 0; j != OUTPUT_LENGTH; ++j)
+      output[j] = get_byte(j % 4, digest[j/4]);
+   }
+
+/*************************************************
+* Clear memory of sensitive data                 *
+*************************************************/
+void SHA_160::clear() throw()
+   {
+   MDx_HashFunction::clear();
+   W.clear();
+   digest[0] = 0x67452301;
+   digest[1] = 0xEFCDAB89;
+   digest[2] = 0x98BADCFE;
+   digest[3] = 0x10325476;
+   digest[4] = 0xC3D2E1F0;
+   }
+
+/*************************************************
+* SHA_160 Constructor                            *
+*************************************************/
+SHA_160::SHA_160() : MDx_HashFunction(20, 64, true, true), W(81)
+   {
+   clear();
+   }
+
+}
diff --git a/src/hash/sha1_ia32/sha1_asm.S b/src/hash/sha1_ia32/sha1_asm.S
new file mode 100644
index 000000000..85bc9dc2c
--- /dev/null
+++ b/src/hash/sha1_ia32/sha1_asm.S
@@ -0,0 +1,242 @@
+/*************************************************
+* SHA-160 Source File                            *
+* (C) 1999-2007 Jack Lloyd                       *
+*************************************************/
+
+#include <botan/asm_macr.h>
+
+START_LISTING(sha1_asm.S)
+
+START_FUNCTION(botan_sha160_asm_ia32)
+   SPILL_REGS()
+
+#define PUSHED 4
+
+   ASSIGN(EDI, ARG(2))
+   ASSIGN(EBP, ARG(3))
+
+   ZEROIZE(ESI)
+
+START_LOOP(.LOAD_INPUT)
+   ADD_IMM(ESI, 4)
+
+   ASSIGN(EAX, ARRAY4(EDI, 0))
+   ASSIGN(EBX, ARRAY4(EDI, 1))
+   ASSIGN(ECX, ARRAY4(EDI, 2))
+   ASSIGN(EDX, ARRAY4(EDI, 3))
+
+   ADD_IMM(EDI, 16)
+
+   BSWAP(EAX)
+   BSWAP(EBX)
+   BSWAP(ECX)
+   BSWAP(EDX)
+
+   ASSIGN(ARRAY4_INDIRECT(EBP,ESI,-4), EAX)
+   ASSIGN(ARRAY4_INDIRECT(EBP,ESI,-3), EBX)
+   ASSIGN(ARRAY4_INDIRECT(EBP,ESI,-2), ECX)
+   ASSIGN(ARRAY4_INDIRECT(EBP,ESI,-1), EDX)
+LOOP_UNTIL_EQ(ESI, 16, .LOAD_INPUT)
+
+   ADD2_IMM(EDI, EBP, 64)
+
+START_LOOP(.EXPANSION)
+   ADD_IMM(ESI, 4)
+
+   ZEROIZE(EAX)
+   ASSIGN(EBX, ARRAY4(EDI, -1))
+   ASSIGN(ECX, ARRAY4(EDI, -2))
+   ASSIGN(EDX, ARRAY4(EDI, -3))
+
+   XOR(EAX, ARRAY4(EDI, -5))
+   XOR(EBX, ARRAY4(EDI, -6))
+   XOR(ECX, ARRAY4(EDI, -7))
+   XOR(EDX, ARRAY4(EDI, -8))
+
+   XOR(EAX, ARRAY4(EDI, -11))
+   XOR(EBX, ARRAY4(EDI, -12))
+   XOR(ECX, ARRAY4(EDI, -13))
+   XOR(EDX, ARRAY4(EDI, -14))
+
+   XOR(EAX, ARRAY4(EDI, -13))
+   XOR(EBX, ARRAY4(EDI, -14))
+   XOR(ECX, ARRAY4(EDI, -15))
+   XOR(EDX, ARRAY4(EDI, -16))
+
+   ROTL_IMM(EDX, 1)
+   ROTL_IMM(ECX, 1)
+   ROTL_IMM(EBX, 1)
+   XOR(EAX, EDX)
+   ROTL_IMM(EAX, 1)
+
+   ASSIGN(ARRAY4(EDI, 0), EDX)
+   ASSIGN(ARRAY4(EDI, 1), ECX)
+   ASSIGN(ARRAY4(EDI, 2), EBX)
+   ASSIGN(ARRAY4(EDI, 3), EAX)
+
+   ADD_IMM(EDI, 16)
+LOOP_UNTIL_EQ(ESI, 80, .EXPANSION)
+
+#define MAGIC1 0x5A827999
+#define MAGIC2 0x6ED9EBA1
+#define MAGIC3 0x8F1BBCDC
+#define MAGIC4 0xCA62C1D6
+
+#define MSG ESP
+#define T2 EBP
+
+#define F1(A, B, C, D, E, F, N)       \
+   ASSIGN(T2, ARRAY4(MSG, N))       ; \
+   ASSIGN(A, F)                     ; \
+   ROTL_IMM(F, 5)                   ; \
+   ADD(F, E)                        ; \
+   ASSIGN(E, C)                     ; \
+   XOR(E, D)                        ; \
+   ADD3_IMM(F, T2, MAGIC1)          ; \
+   AND(E, B)                        ; \
+   XOR(E, D)                        ; \
+   ROTR_IMM(B, 2)                   ; \
+   ADD(E, F)                        ;
+
+#define F2_4(A, B, C, D, E, F, N, MAGIC) \
+   ASSIGN(T2, ARRAY4(MSG, N))       ; \
+   ASSIGN(A, F)                     ; \
+   ROTL_IMM(F, 5)                   ; \
+   ADD(F, E)                        ; \
+   ASSIGN(E, B)                     ; \
+   XOR(E, C)                        ; \
+   ADD3_IMM(F, T2, MAGIC)           ; \
+   XOR(E, D)                        ; \
+   ROTR_IMM(B, 2)                   ; \
+   ADD(E, F)                        ;
+
+#define F3(A, B, C, D, E, F, N)       \
+   ASSIGN(T2, ARRAY4(MSG, N))       ; \
+   ASSIGN(A, F)                     ; \
+   ROTL_IMM(F, 5)                   ; \
+   ADD(F, E)                        ; \
+   ASSIGN(E, B)                     ; \
+   OR(E, C)                         ; \
+   AND(E, D)                        ; \
+   ADD3_IMM(F, T2, MAGIC3)          ; \
+   ASSIGN(T2, B)                    ; \
+   AND(T2, C)                       ; \
+   OR(E, T2)                        ; \
+   ROTR_IMM(B, 2)                   ; \
+   ADD(E, F)                        ;
+
+#define F2(A, B, C, D, E, F, MSG) \
+   F2_4(A, B, C, D, E, F, MSG, MAGIC2)
+
+#define F4(A, B, C, D, E, F, MSG) \
+   F2_4(A, B, C, D, E, F, MSG, MAGIC4)
+
+   ASSIGN(EAX, ARG(1))
+   ASSIGN(EDI, ARRAY4(EAX, 0))
+   ASSIGN(EBX, ARRAY4(EAX, 1))
+   ASSIGN(ECX, ARRAY4(EAX, 2))
+   ASSIGN(EDX, ARRAY4(EAX, 3))
+   ASSIGN(ESI, ARRAY4(EAX, 4))
+
+   ASSIGN(ARRAY4(EBP, 80), ESP)
+   ASSIGN(ESP, EBP)
+
+   /* First Round */
+   F1(EAX, EBX, ECX, EDX, ESI, EDI, 0)
+   F1(EDI, EAX, EBX, ECX, EDX, ESI, 1)
+   F1(ESI, EDI, EAX, EBX, ECX, EDX, 2)
+   F1(EDX, ESI, EDI, EAX, EBX, ECX, 3)
+   F1(ECX, EDX, ESI, EDI, EAX, EBX, 4)
+   F1(EBX, ECX, EDX, ESI, EDI, EAX, 5)
+   F1(EAX, EBX, ECX, EDX, ESI, EDI, 6)
+   F1(EDI, EAX, EBX, ECX, EDX, ESI, 7)
+   F1(ESI, EDI, EAX, EBX, ECX, EDX, 8)
+   F1(EDX, ESI, EDI, EAX, EBX, ECX, 9)
+   F1(ECX, EDX, ESI, EDI, EAX, EBX, 10)
+   F1(EBX, ECX, EDX, ESI, EDI, EAX, 11)
+   F1(EAX, EBX, ECX, EDX, ESI, EDI, 12)
+   F1(EDI, EAX, EBX, ECX, EDX, ESI, 13)
+   F1(ESI, EDI, EAX, EBX, ECX, EDX, 14)
+   F1(EDX, ESI, EDI, EAX, EBX, ECX, 15)
+   F1(ECX, EDX, ESI, EDI, EAX, EBX, 16)
+   F1(EBX, ECX, EDX, ESI, EDI, EAX, 17)
+   F1(EAX, EBX, ECX, EDX, ESI, EDI, 18)
+   F1(EDI, EAX, EBX, ECX, EDX, ESI, 19)
+
+   /* Second Round */
+   F2(ESI, EDI, EAX, EBX, ECX, EDX, 20)
+   F2(EDX, ESI, EDI, EAX, EBX, ECX, 21)
+   F2(ECX, EDX, ESI, EDI, EAX, EBX, 22)
+   F2(EBX, ECX, EDX, ESI, EDI, EAX, 23)
+   F2(EAX, EBX, ECX, EDX, ESI, EDI, 24)
+   F2(EDI, EAX, EBX, ECX, EDX, ESI, 25)
+   F2(ESI, EDI, EAX, EBX, ECX, EDX, 26)
+   F2(EDX, ESI, EDI, EAX, EBX, ECX, 27)
+   F2(ECX, EDX, ESI, EDI, EAX, EBX, 28)
+   F2(EBX, ECX, EDX, ESI, EDI, EAX, 29)
+   F2(EAX, EBX, ECX, EDX, ESI, EDI, 30)
+   F2(EDI, EAX, EBX, ECX, EDX, ESI, 31)
+   F2(ESI, EDI, EAX, EBX, ECX, EDX, 32)
+   F2(EDX, ESI, EDI, EAX, EBX, ECX, 33)
+   F2(ECX, EDX, ESI, EDI, EAX, EBX, 34)
+   F2(EBX, ECX, EDX, ESI, EDI, EAX, 35)
+   F2(EAX, EBX, ECX, EDX, ESI, EDI, 36)
+   F2(EDI, EAX, EBX, ECX, EDX, ESI, 37)
+   F2(ESI, EDI, EAX, EBX, ECX, EDX, 38)
+   F2(EDX, ESI, EDI, EAX, EBX, ECX, 39)
+
+   /* Third Round */
+   F3(ECX, EDX, ESI, EDI, EAX, EBX, 40)
+   F3(EBX, ECX, EDX, ESI, EDI, EAX, 41)
+   F3(EAX, EBX, ECX, EDX, ESI, EDI, 42)
+   F3(EDI, EAX, EBX, ECX, EDX, ESI, 43)
+   F3(ESI, EDI, EAX, EBX, ECX, EDX, 44)
+   F3(EDX, ESI, EDI, EAX, EBX, ECX, 45)
+   F3(ECX, EDX, ESI, EDI, EAX, EBX, 46)
+   F3(EBX, ECX, EDX, ESI, EDI, EAX, 47)
+   F3(EAX, EBX, ECX, EDX, ESI, EDI, 48)
+   F3(EDI, EAX, EBX, ECX, EDX, ESI, 49)
+   F3(ESI, EDI, EAX, EBX, ECX, EDX, 50)
+   F3(EDX, ESI, EDI, EAX, EBX, ECX, 51)
+   F3(ECX, EDX, ESI, EDI, EAX, EBX, 52)
+   F3(EBX, ECX, EDX, ESI, EDI, EAX, 53)
+   F3(EAX, EBX, ECX, EDX, ESI, EDI, 54)
+   F3(EDI, EAX, EBX, ECX, EDX, ESI, 55)
+   F3(ESI, EDI, EAX, EBX, ECX, EDX, 56)
+   F3(EDX, ESI, EDI, EAX, EBX, ECX, 57)
+   F3(ECX, EDX, ESI, EDI, EAX, EBX, 58)
+   F3(EBX, ECX, EDX, ESI, EDI, EAX, 59)
+
+   /* Fourth Round */
+   F4(EAX, EBX, ECX, EDX, ESI, EDI, 60)
+   F4(EDI, EAX, EBX, ECX, EDX, ESI, 61)
+   F4(ESI, EDI, EAX, EBX, ECX, EDX, 62)
+   F4(EDX, ESI, EDI, EAX, EBX, ECX, 63)
+   F4(ECX, EDX, ESI, EDI, EAX, EBX, 64)
+   F4(EBX, ECX, EDX, ESI, EDI, EAX, 65)
+   F4(EAX, EBX, ECX, EDX, ESI, EDI, 66)
+   F4(EDI, EAX, EBX, ECX, EDX, ESI, 67)
+   F4(ESI, EDI, EAX, EBX, ECX, EDX, 68)
+   F4(EDX, ESI, EDI, EAX, EBX, ECX, 69)
+   F4(ECX, EDX, ESI, EDI, EAX, EBX, 70)
+   F4(EBX, ECX, EDX, ESI, EDI, EAX, 71)
+   F4(EAX, EBX, ECX, EDX, ESI, EDI, 72)
+   F4(EDI, EAX, EBX, ECX, EDX, ESI, 73)
+   F4(ESI, EDI, EAX, EBX, ECX, EDX, 74)
+   F4(EDX, ESI, EDI, EAX, EBX, ECX, 75)
+   F4(ECX, EDX, ESI, EDI, EAX, EBX, 76)
+   F4(EBX, ECX, EDX, ESI, EDI, EAX, 77)
+   F4(EAX, EBX, ECX, EDX, ESI, EDI, 78)
+   F4(EDI, EAX, EBX, ECX, EDX, ESI, 79)
+
+   ASSIGN(ESP, ARRAY4(ESP, 80))
+
+   ASSIGN(EBP, ARG(1))
+   ADD(ARRAY4(EBP, 0), EDX)
+   ADD(ARRAY4(EBP, 1), EDI)
+   ADD(ARRAY4(EBP, 2), EAX)
+   ADD(ARRAY4(EBP, 3), EBX)
+   ADD(ARRAY4(EBP, 4), ECX)
+
+   RESTORE_REGS()
+END_FUNCTION(botan_sha160_asm_ia32)