[keccak-asm] keccak_f1600 in asm

diff --git a/keccak/keccak-asm.S b/keccak/keccak-asm.S
index 77dc9ce10bc99f92f67f5b7e43ddcf0155c5532c..357c8bd028fd36d6a1baa56358c42978ff22157c 100644 (file)
--- a/keccak/keccak-asm.S
+++ b/keccak/keccak-asm.S
@@ -166,14 +166,18 @@ chi_step:
        brne 10b
        ret
 
-.global keccak_theta
-keccak_theta:
-       push_range 2, 8
+.global keccak_f1600
+keccak_f1600:
+       push_range 2, 9
        push r16
        push_range 28, 29
 
+       stack_alloc_large 200, r26, r27
+       adiw XL, 1
+
+       clr r9
+5:
        movw r30, r24 ; Z = a
-       movw r26, r22 ; X = b
 
        ldi r19, 5
 10:
@@ -350,6 +354,7 @@ keccak_theta:
 
        ; Z points at b
        movw XL, ZL
+       movw r4, ZL
        adiw XL, 8
        adiw ZL, 16
        movw YL, r2
@@ -366,8 +371,51 @@ keccak_theta:
        adiw ZL, 5 * 8
        dec r18
        brne 10b
+
+       /* -- iota -- */
+       ldi r30, lo8(keccak_rc_comp)
+       ldi r31, hi8(keccak_rc_comp)
+       add r30, r9
+       adc r31, __zero_reg__
+       lpm r20, Z+
+       movw YL, r2
+       ldi r21, 0x80
+       bst r20, 6
+       brtc 10f
+       ldd r22, Y+7
+       eor r22, r21
+       std Y+7, r22
+10:
+       bst r20, 5
+       brtc 10f
+       ldd r22, Y+3
+       eor r22, r21
+       std Y+3, r22
+10:
+       bst r20, 4
+       brtc 10f
+       ldd r22, Y+1
+       eor r22, r21
+       std Y+1, r22
+10:
+       andi r20, 0x8f
+       ld r22, Y
+       eor r22, r20
+       st Y, r22
+
+       inc r9
+       mov r16, r9
+       cpi r16, 24
+       breq 20f
+       movw r24, YL
+       movw r26, r4
+       rjmp 5b
+20:
+
+       stack_free_large3 200
+
        pop_range 28, 29
        pop r16
-       pop_range 2, 8
+       pop_range 2, 9
 
        ret

diff --git a/keccak/keccak-stub.c b/keccak/keccak-stub.c
index 5927614e001870d833377a4153c9f8f50f5adc54..4e5a58abe7cce5569a1527eaf9b1a182f9064be0 100644 (file)
--- a/keccak/keccak-stub.c
+++ b/keccak/keccak-stub.c
@@ -79,7 +79,7 @@ const uint64_t rc[] PROGMEM = {
 };
 */
 
-const static uint8_t rc_comp[] PROGMEM = {
+const uint8_t keccak_rc_comp[] PROGMEM = {
                0x01, 0x92, 0xda, 0x70,
                0x9b, 0x21, 0xf1, 0x59,
                0x8a, 0x88, 0x39, 0x2a,
@@ -104,96 +104,7 @@ const uint8_t keccak_rotate_codes[5][5] PROGMEM = {
         { ROT_CODE(18), ROT_CODE( 2), ROT_CODE(61), ROT_CODE(56), ROT_CODE(14) }
 };
 
-void keccak_theta(uint64_t *a, uint64_t *b);
-extern const uint8_t rho_pi_idx_table[25] PROGMEM;
-
-static inline
-void keccak_round(uint64_t a[5][5], uint8_t rci){
-       uint64_t b[5][5];
-//     uint8_t i, j;
-       union {
-                       uint64_t v64;
-                       uint8_t v8[8];
-               } t;
-       /* theta */
-       keccak_theta((uint64_t*)a, (uint64_t*)b);
-#if DEBUG
-       cli_putstr_P(PSTR("\r\nAfter theta:"));
-       keccak_dump_state(a);
-#endif
-       /* rho & pi */
-/*
-       const uint8_t* rot_code = (const uint8_t*)keccak_rotate_codes;
-    const uint8_t* idx_idx = (const uint8_t*)rho_pi_idx_table;
-    uint64_t *a_tmp = (uint64_t*)a;
-       for(i = 0; i < 25; ++i){
-                   *((uint64_t*)(((uint8_t*)b) + pgm_read_byte(idx_idx++))) =
-                rotate64left_code(*a_tmp++, pgm_read_byte(rot_code++));
-
-       }
-*/
-#if DEBUG & 0
-       cli_putstr_P(PSTR("\r\n--- after rho & pi ---"));
-       keccak_dump_state(a);
-#endif
-       /* chi */
-//     memcpy(a, b, 5 * 5 * 8);
-//     for(i = 1; i < 5; ++i){
-/*
-               for(j = 0; j < 5; ++j){
-                       a[i][j] =  b[i][j] ^ ((~(b[i][(j + 1) % 5])) & (b[i][(j + 2) % 5]));
-               }
-* /
-//           a[i][0] ^= ((~(b[i][1])) & (b[i][2]));
-//        a[i][1] ^= ((~(b[i][2])) & (b[i][3]));
-//        a[i][2] ^= ((~(b[i][3])) & (b[i][4]));
-        for(j = 0; j < 3 * 8; ++j){
-            ((uint8_t*)a)[i * 5 * 8 + j] ^=
-                (~((uint8_t*)b)[i * 5 * 8 + j + 8]) & ((uint8_t*)a)[i * 5 * 8 + j + 16];
-        }
-        a[i][3] ^= ((~(b[i][4])) & (b[i][0]));
-        a[i][4] ^= ((~(b[i][0])) & (b[i][1]));
-
-       }
-*/
-#if DEBUG & 0
-       cli_putstr_P(PSTR("\r\nAfter chi:"));
-       keccak_dump_state(a);
-#endif
-       /* iota */
-
-//     memcpy_P(&t, &(rc_comp[rci]), 8);
-       t.v64 = 0;
-       t.v8[0] = pgm_read_byte(&(rc_comp[rci]));
-       if(t.v8[0] & 0x40){
-               t.v8[7] = 0x80;
-       }
-       if(t.v8[0] & 0x20){
-               t.v8[3] = 0x80;
-       }
-       if(t.v8[0] & 0x10){
-               t.v8[1] = 0x80;
-       }
-       t.v8[0] &= 0x8F;
-
-       a[0][0] ^= t.v64;
-#if DEBUG & 0
-       cli_putstr_P(PSTR("\r\nAfter iota:"));
-       keccak_dump_state(a);
-#endif
-}
-
-void keccak_f1600(uint64_t a[5][5]){
-       uint8_t i = 0;
-       do {
-#if DEBUG
-               cli_putstr_P(PSTR("\r\n\r\n--- Round "));
-               cli_hexdump(&i, 1);
-               cli_putstr_P(PSTR(" ---"));
-#endif
-               keccak_round(a, i);
-       } while (++i < 24);
-}
+void keccak_f1600(uint64_t a[5][5]);
 
 void keccak_nextBlock(keccak_ctx_t* ctx, const void* block){
        memxor(ctx->a, block, ctx->bs);