even smaller BMW224/256 (1988 bytes)

diff --git a/bmw/bmw_small-tinyasm.S b/bmw/bmw_small-tinyasm.S
index d0f8530ffd160099578572974021e8aa8a591975..198aa7024688413f0bcb985079237e80bb3532b1 100644 (file)
--- a/bmw/bmw_small-tinyasm.S
+++ b/bmw/bmw_small-tinyasm.S
@@ -122,41 +122,45 @@ eor_r22_in_r16:
   param s: r20
 */
 sn:
-       push_range 12, 20
+       push_range 2, 5
+       push r17
+       push r19
        ldi r30, lo8(s_table)
        ldi r31, hi8(s_table)
        lsl r20
        lsl r20
        add r30, r20
        adc r31, r1
-       movw r12, r22
-       movw r14, r24
+       movw r2, r22
+       movw r4, r24
        lpm r20, Z+
        rcall shiftright32
        movw r16, r22
        movw r18, r24
 ;---
-       movw r22, r12
-       movw r24, r14
+       movw r22, r2
+       movw r24, r4
        lpm r20, Z+
        rcall shiftleft32
        rcall eor_r22_in_r16
 ;---
-       movw r22, r12
-       movw r24, r14
+       movw r22, r2
+       movw r24, r4
        lpm r20, Z+
        rcall rotateleft32
        rcall eor_r22_in_r16
 ;---
-       movw r22, r12
-       movw r24, r14
+       movw r22, r2
+       movw r24, r4
        lpm r20, Z+
        rcall rotateleft32
        eor r22, r16
        eor r23, r17
        eor r24, r18
        eor r25, r19
-       pop_range 12, 20
+       pop r19
+       pop r17
+       pop_range 2, 5
        ret
 
 /******************************************************************************/
@@ -236,6 +240,17 @@ load32_from_Y:
        ld r24, Y+
        ld r25, Y+
        ret
+
+add_X_to_32:
+       ld r0, X+
+       add r22, r0
+       ld r0, X+
+       adc r23, r0
+       ld r0, X+
+       adc r24, r0
+       ld r0, X+
+       adc r25, r0
+       ret
 /******************************************************************************/
 /*
   param q:  r28:r29 (Y)
@@ -345,14 +360,7 @@ f0:
        rcall sn
        movw r30, r2
 
-       ld r0, X+
-       add r22, r0
-       ld r0, X+
-       adc r23, r0
-       ld r0, X+
-       adc r24, r0
-       ld r0, X+
-       adc r25, r0
+       rcall add_X_to_32
 
        st Y+, r22
        st Y+, r23
@@ -369,19 +377,15 @@ f0:
        rcall sn
        movw r30, r2
        movw r26, h0
-       ld r0, X+
-       add r22, r0
-       ld r0, X+
-       adc r23, r0
-       ld r0, X+
-       adc r24, r0
-       ld r0, X+
-       adc r25, r0
+       rcall add_X_to_32
        sbiw r26, 4
-       st Y+, r22
-       st Y+, r23
-       st Y+, r24
-       st Y+, r25
+       std Y+0, r22
+       std Y+1, r23
+       std Y+2, r24
+       std Y+3, r25
+       sbiw r28, 15*4
+       movw r20, h0
+       movw r22, m0
        ret
 
 /******************************************************************************/
@@ -627,6 +631,9 @@ f1:
        inc r17
        sbrs r17, 4
        rjmp 10b
+       movw r24, r2
+       movw r22, r4
+       movw r20, r6
        ret
 
 /******************************************************************************/
@@ -860,8 +867,8 @@ bmw256_nextBlock:
        push_range  2, 17
        stack_alloc_large 32*4, r28, r29
        adiw r28, 1
-       push_range 28, 29 /* push Q */
-       push_range 22, 25 /* push M & H */
+;      push_range 28, 29 /* push Q */
+;      push_range 22, 25 /* push M & H */
        /* increment counter */
        movw r26, r24
        movw r2, r26
@@ -882,20 +889,14 @@ bmw256_nextBlock:
        movw r26, r24
        rcall f0
        /* call f1*/
-       pop r21
-       pop r20
-       pop r23
-       pop r22
-       pop r25
-       pop r24
+       movw r24, r28
+
 ;      rcall printQ
-       push_range 20, 25
        rcall f1
        /* call f2 */
 ;      pop_range 20, 25
 ;      push_range 20, 25
 ;      rcall printQ
-       pop_range 20, 25
 ;      push r20
 ;      push r21
        call f2