further shrinked bmw tiny to 1714 bytes

diff --git a/bmw/bmw_small-tinyasm.S b/bmw/bmw_small-tinyasm.S
index 7cdbadb59c4ebe93093a5bd09e594aee1c0405a2..9b7b2a570202de0e8f26446a7b3caa7258424834 100644 (file)
--- a/bmw/bmw_small-tinyasm.S
+++ b/bmw/bmw_small-tinyasm.S
@@ -28,6 +28,11 @@
 
 #include "avr-asm-macros.S"
 
+acc2 =  8
+acc3 =  9
+acc0 = 14
+acc1 = 15
+
 /******************************************************************************/
 /*
   param a: r22:r23:r24:r25
@@ -91,6 +96,7 @@ bitrotateleft_1:
        breq 20f
 10:
        lsl r0
+rol32:
        rol r22
        rol r23
        rol r24
@@ -102,6 +108,18 @@ bitrotateleft_1:
 
 /******************************************************************************/
 
+sn_stub:
+       movw r22, r2
+       movw r24, r4
+       lpm r20, Z+
+       rcall rotateleft32
+eor32_to_acc:
+       eor acc0, r22
+       eor acc1, r23
+       eor acc2, r24
+       eor acc3, r25
+       ret
+
 s_table:
 s0:  .byte 1, 3, 4,19
 s1:  .byte 1, 2, 8,23
@@ -117,14 +135,10 @@ s3:  .byte 0xAF, 29
 s4:  .byte 0x00,  0
 s5:  .byte 0x80,  0
 */
-acc2 =  8
-acc3 =  9
 h0   = 10
 h1   = 11
 m0   = 12
 m1   = 13
-acc0 = 14
-acc1 = 15
 
 /*
   param x: r22:r23:r24:25
@@ -155,17 +169,9 @@ sn:
        rcall shiftleft32
        rcall eor32_to_acc
 ;---
-       movw r22, r2
-       movw r24, r4
-       lpm r20, Z+
-       rcall rotateleft32
-       rcall eor32_to_acc
-;---
-       movw r22, r2
-       movw r24, r4
-       lpm r20, Z+
-       rcall rotateleft32
-       rcall eor32_to_acc
+       rcall sn_stub
+       rcall sn_stub
+
        movw r22, acc0
        movw r24, acc2
        pop acc3
@@ -236,13 +242,6 @@ add_X_to_32:
        adc r25, r0
        ret
 
-store_acc_to_dec_X:
-       st -X, acc3
-       st -X, acc2
-       st -X, acc1
-       st -X, acc0
-       ret
-
 store32_to_X:
        st X+, r22
        st X+, r23
@@ -290,20 +289,6 @@ m1   = 13
 acc0 = 14
 acc1 = 15
 
-add32_to_acc:
-       add acc0, r22
-       adc acc1, r23
-       adc acc2, r24
-       adc acc3, r25
-       ret
-
-eor32_to_acc:
-       eor acc0, r22
-       eor acc1, r23
-       eor acc2, r24
-       eor acc3, r25
-       ret
-
 load_acc_from_X:
        ld acc0, X+
        ld acc1, X+
@@ -327,6 +312,7 @@ add_acc_to_Z:
        ret
 
 load_rotate_add_M:
+       mov r20, j
        andi r20, 0x0f
        mov r0, r20
        lsl r0
@@ -361,33 +347,33 @@ addelement:
        rcall store_acc_to_dec_X
        adiw r26, 4
        clt
-       mov r20, j
        rcall load_rotate_add_M
-       mov r20, j
-       subi r20, -3
+       subi j, -3
        rcall load_rotate_add_M
-       mov r20, j
        set
-       subi r20, -10
+       subi j, -7
        rcall load_rotate_add_M
        lsl j
        lsl j
-       subi j, -7*4
+       subi j, -7*4+10*4
        andi j, 0x3f
        movw r26, h0
        add r26, j
        adc r27, r1
        rcall load32_from_X
-       rcall eor32_to_acc
+       rjmp eor32_to_acc
 ;---
-       ret
 
 /******************************************************************************/
 load_sn_add:
        rcall load32_from_X
        rcall sn
-       rjmp add32_to_acc
-;      ret
+add32_to_acc:
+       add acc0, r22
+       adc acc1, r23
+       adc acc2, r24
+       adc acc3, r25
+       ret
 
 /*
   param q: r26:r27
@@ -458,8 +444,12 @@ expand2:
        rcall load_sn_add
 expand2_exit:
        adiw r26, 4
-       rjmp store_acc_to_dec_X
-;      ret
+store_acc_to_dec_X:
+       st -X, acc3
+       st -X, acc2
+       st -X, acc1
+       st -X, acc0
+       ret
 
 /******************************************************************************/
 /*
@@ -998,11 +988,9 @@ bmw256_lastBlock:
        rcall load32_from_Z_stub
 410:
        clr r25
+       ldi r20, 1
        lsl r21
-       rol r22
-       rol r23
-       rol r24
-       rol r25
+       rcall rol32
        mov r20, len0
        add r21, len1
        adc r22, r1
@@ -1068,10 +1056,9 @@ bmw256_lastBlock:
 */
 .global bmw224_ctx2hash
 bmw224_ctx2hash:
-       movw r26, r24
        movw r30, r22
        adiw r30, 9*4
-       ldi r22, 28
+       ldi r18, 28
        rjmp 1f
 
 /*******************************************************************************
@@ -1084,14 +1071,13 @@ bmw224_ctx2hash:
 */
 .global bmw256_ctx2hash
 bmw256_ctx2hash:
-       movw r26, r24
        movw r30, r22
        adiw r30, 8*4
-       ldi r22, 32
-1:
-       ld r23, Z+
+       ldi r18, 32
+1:     movw r26, r24
+1:  ld r23, Z+
        st X+, r23
-       dec r22
+       dec r18
        brne 1b
        ret
 
@@ -1229,36 +1215,30 @@ c2h_lut:
 */
 .global bmw224_init
 bmw224_init:
-       movw r26, r24
-       ldi r22, 0x03
-       ldi r23, 0x02
-       ldi r24, 0x01
-       ldi r25, 0x00
+       ldi r22, 0x00
+       ldi r23, 0x40
 bmw_small_init:
-       rcall store32_to_X
-       ldi r18, 16-1
-       ldi r20, 0x04
-1:
-       add r22, r20
-       adc r23, r20
-       adc r24, r20
-       adc r25, r20
-       rcall store32_to_X
-       dec r18
-       brne 1b
-       st X+, r1
-       st X+, r1
-       st X+, r1
-       st X+, r1
+       movw r26, r24
+       adiw r26, 4
+10:
+       st -X, r22
+       inc r22
+       mov r20, r22
+       andi r20, 0x3
+       brne 10b
+       adiw r26, 8
+20: cp r22, r23
+       brne 10b
+       st -X, r1
+       st -X, r1
+       st -X, r1
+       st -X, r1
        ret
 
 .global bmw256_init
 bmw256_init:
-       movw r26, r24
-       ldi r22, 0x43
-       ldi r23, 0x42
-       ldi r24, 0x41
-       ldi r25, 0x40
+       ldi r22, 0x40
+       ldi r23, 0x80
        rjmp bmw_small_init