[keccak-asm] *_ctx2hash in asm

[avr-crypto-lib.git] / keccak / keccak-asm.S

diff --git a/keccak/keccak-asm.S b/keccak/keccak-asm.S
index cec7900bff80a94bd4fb13dc50ecc65301d4e90a..f1fc07a10e561159ec9fd74b3b63c4cc139c18b0 100644 (file)
--- a/keccak/keccak-asm.S
+++ b/keccak/keccak-asm.S
@@ -43,10 +43,6 @@ ctx_a:
        .struct ctx_a + 8 * 5 * 5
 ctx_r:
        .struct ctx_r + 2
-ctx_c:
-       .struct ctx_c + 2
-ctx_d:
-       .struct ctx_d + 1
 ctx_bs:
 
        .section .text
@@ -544,15 +540,17 @@ icall_r16_times:
 */
        ; memcpy(a, b, 200)
        ; X points at b + 32 + 8 = b + 40 = b[1][0] has to point to b[0][0]
-       ldi r16, 200
+       ldi r16, 200 / 8
        sbiw XL, 5 * 8
        movw ZL, XL
        subi YL, lo8(5 * 5 * 8)
        sbci YH, hi8(5 * 5 * 8)
        movw r2, YL
 10:
+       .rept 8
        ld r22, X+
        st Y+, r22
+       .endr
        dec r16
        brne 10b
 
@@ -624,6 +622,42 @@ icall_r16_times:
        ret
        .endfunc
 
+
+       .global keccak224_ctx2hash
+       .func keccak224_ctx2hash
+keccak224_ctx2hash:
+       movw r20, r22
+       ldi r22, lo8(224)
+       ldi r23, hi8(224)
+       rjmp keccak_ctx2hash
+       .endfunc
+
+       .global keccak384_ctx2hash
+       .func keccak384_ctx2hash
+keccak384_ctx2hash:
+       movw r20, r22
+       ldi r22, lo8(384)
+       ldi r23, hi8(384)
+       rjmp keccak_ctx2hash
+       .endfunc
+
+       .global keccak512_ctx2hash
+       .func keccak512_ctx2hash
+keccak512_ctx2hash:
+       movw r20, r22
+       ldi r22, lo8(512)
+       ldi r23, hi8(512)
+       rjmp keccak_ctx2hash
+       .endfunc
+
+       .global keccak256_ctx2hash
+       .func keccak256_ctx2hash
+keccak256_ctx2hash:
+       movw r20, r22
+       ldi r22, lo8(256)
+       ldi r23, hi8(256)
+       .endfunc
+
 /*
 void keccak_ctx2hash(void* dest, uint16_t length_b, keccak_ctx_t* ctx){
        while(length_b>=ctx->r){
@@ -641,17 +675,19 @@ keccak_ctx2hash:
        push_range 2, 10
        movw r4, r20
        movw r6, r24
-       movw ZL, r24
+       movw ZL, r20
        movw r8, r22
        subi ZL, lo8(-ctx_r)
-       subi ZH, hi8(-ctx_r)
+       sbci ZH, hi8(-ctx_r)
        ld r2, Z+
        ld r3, Z+
        ldd r10, Z+3 ; load blocksize (in bytes)
 10:
-       cp  r8, r2
-       cpc r9, r3
-       brcc 40f
+       ; length_b = (r9:r8) ; r = (r3:r2) ; (H:L)
+       cp  r2, r8
+       cpc r3, r9
+;      rjmp 40f
+       brsh 40f
        movw XL, r4
        movw ZL, r6
        mov r24, r10