now BMW-small entirely in asm

[avr-crypto-lib.git] / bmw / bmw_small-asm.S

diff --git a/bmw/bmw_small-asm.S b/bmw/bmw_small-asm.S
index 1aab11a1cc87aa4f025b27a1fd2540823d801ab2..03f87028d49e6eaee8ed80a4286e7a7a7eff6595 100644 (file)
--- a/bmw/bmw_small-asm.S
+++ b/bmw/bmw_small-asm.S
@@ -2077,3 +2077,136 @@ bmw256_ctx2hash:
        dec r22
        brne 1b
        ret
+
+/*******************************************************************************
+* void bmw256(void* dest, const void* msg, uint32_t length_b){
+*      bmw_small_ctx_t ctx;
+*      bmw256_init(&ctx);
+*      while(length_b>=BMW_SMALL_BLOCKSIZE){
+*              bmw_small_nextBlock(&ctx, msg);
+*              length_b -= BMW_SMALL_BLOCKSIZE;
+*              msg = (uint8_t*)msg + BMW_SMALL_BLOCKSIZE_B;
+*      }
+*      bmw_small_lastBlock(&ctx, msg, length_b);
+*      bmw256_ctx2hash(dest, &ctx);
+* }
+*
+* param dest:     r24:r25
+* param msg:      r22:r23
+* param length_b: r18:r21
+*/
+ctx0 =   2
+ctx1 =   3
+msg0 =   4
+msg1 =   5
+len0 =   6
+len1 =   7
+len2 =   8
+len3 =   9
+dst0 =  10
+dst1 =  11
+.global bmw256
+bmw256:
+       push_range 2, 11
+       stack_alloc_large 64+4
+       adiw r30, 1
+       movw ctx0, r30
+       movw dst0, r24
+       movw msg0, r22
+       movw len0, r18
+       movw len2, r20
+       movw r24, ctx0
+       rcall bmw256_init
+20:
+       mov r18, len2
+       or  r18, len3
+       breq 50f
+       movw r24, ctx0
+       movw r22, msg0
+       rcall bmw_small_nextBlock
+       ldi r20, 2
+       sub len1, r20
+       sbc len2, r1
+       sbc len3, r1
+       ldi r20, 64
+       add msg0, r20
+       adc msg1, r1
+       rjmp 20b
+50:
+       movw r24, ctx0
+       movw r22, msg0
+       movw r20, len0
+       rcall bmw_small_lastBlock
+       movw r24, dst0
+       movw r22, ctx0
+       rcall bmw256_ctx2hash
+       stack_free_large 64+4
+       pop_range 2, 11
+       ret
+
+
+/*******************************************************************************
+* void bmw224(void* dest, const void* msg, uint32_t length_b){
+*      bmw_small_ctx_t ctx;
+*      bmw224_init(&ctx);
+*      while(length_b>=BMW_SMALL_BLOCKSIZE){
+*              bmw_small_nextBlock(&ctx, msg);
+*              length_b -= BMW_SMALL_BLOCKSIZE;
+*              msg = (uint8_t*)msg + BMW_SMALL_BLOCKSIZE_B;
+*      }
+*      bmw_small_lastBlock(&ctx, msg, length_b);
+*      bmw224_ctx2hash(dest, &ctx);
+* }
+*
+* param dest:     r24:r25
+* param msg:      r22:r23
+* param length_b: r18:r21
+*/
+ctx0 =   2
+ctx1 =   3
+msg0 =   4
+msg1 =   5
+len0 =   6
+len1 =   7
+len2 =   8
+len3 =   9
+dst0 =  10
+dst1 =  11
+.global bmw224
+bmw224:
+       push_range 2, 11
+       stack_alloc_large 64+4
+       adiw r30, 1
+       movw ctx0, r30
+       movw dst0, r24
+       movw msg0, r22
+       movw len0, r18
+       movw len2, r20
+       movw r24, ctx0
+       rcall bmw224_init
+20:
+       mov r18, len2
+       or  r18, len3
+       breq 50f
+       movw r24, ctx0
+       movw r22, msg0
+       rcall bmw_small_nextBlock
+       ldi r20, 2
+       sub len1, r20
+       sbc len2, r1
+       sbc len3, r1
+       ldi r20, 64
+       add msg0, r20
+       adc msg1, r1
+       rjmp 20b
+50:
+       movw r24, ctx0
+       movw r22, msg0
+       movw r20, len0
+       rcall bmw_small_lastBlock
+       movw r24, dst0
+       movw r22, ctx0
+       rcall bmw224_ctx2hash
+       stack_free_large 64+4
+       pop_range 2, 11
+       ret