further shrinked bmw tiny to 1694 bytes

diff --git a/bmw/bmw_small-tinyasm.S b/bmw/bmw_small-tinyasm.S
index 9b7b2a570202de0e8f26446a7b3caa7258424834..94f1825a7b50704351877249f826f7aeff9ac760 100644 (file)
--- a/bmw/bmw_small-tinyasm.S
+++ b/bmw/bmw_small-tinyasm.S
@@ -160,8 +160,7 @@ sn:
        movw r4, r24
        lpm r20, Z+
        rcall shiftright32
-       movw acc0, r22
-       movw acc2, r24
+       rcall mov32_to_acc
 ;---
        movw r22, r2
        movw r24, r4
@@ -249,6 +248,16 @@ store32_to_X:
        st X+, r25
        ret
 
+mov32_to_acc:
+       movw acc0, r22
+       movw acc2, r24
+       ret
+
+eor_acc_from_Y_add_to_Z:
+       rcall load32_from_Y
+       rcall eor32_to_acc
+       rjmp  add_acc_to_Z
+
 /******************************************************************************/
 /*
   param q:  r28:r29 (Y)
@@ -332,6 +341,30 @@ load_rotate_add_M:
        sbc acc3, r25
        ret
 
+
+;---
+
+/******************************************************************************/
+load_sn_add:
+       rcall load32_from_X
+       rcall sn
+add32_to_acc:
+       add acc0, r22
+       adc acc1, r23
+       adc acc2, r24
+       adc acc3, r25
+       ret
+
+/*
+  param q: r26:r27
+  param m: r22:r23
+  param h: r20:r21
+  param j: r24
+*/
+
+expand_intro:
+       push_range 26, 27
+       push r24
 addelement:
        mov j, r24
        movw h0, r20
@@ -361,32 +394,8 @@ addelement:
        add r26, j
        adc r27, r1
        rcall load32_from_X
-       rjmp eor32_to_acc
-;---
-
-/******************************************************************************/
-load_sn_add:
-       rcall load32_from_X
-       rcall sn
-add32_to_acc:
-       add acc0, r22
-       adc acc1, r23
-       adc acc2, r24
-       adc acc3, r25
-       ret
-
-/*
-  param q: r26:r27
-  param m: r22:r23
-  param h: r20:r21
-  param j: r24
-*/
-
-expand_intro:
-
-       push_range 26, 27
-       push r24
-       rcall addelement
+       rcall eor32_to_acc
+;--
        pop r24
        pop_range 26, 27
        lsl r24
@@ -498,6 +507,11 @@ h1 = 5
 m0 = 6
 m1 = 7
 
+restore_f1:
+       movw r26, r2
+       movw r22, r4
+    movw r20, r6
+       ret
 
 .global bmw_small_nextBlock
 .global bmw224_nextBlock
@@ -644,23 +658,18 @@ f1:
        movw r26, r2
        clr r24
        rcall expand1
-       movw r26, r2
-       movw r22, r4
-    movw r20, r6
+       rcall restore_f1
        ldi r24, 1
        rcall expand1
        ldi r17, 2
-10:    movw r26, r2
-       movw r22, r4
-       movw r20, r6
+10: rcall restore_f1
        mov r24, r17
        rcall expand2
        inc r17
        sbrs r17, 4
        rjmp 10b
+       rcall restore_f1
        movw r24, r2
-       movw r22, r4
-       movw r20, r6
 
 
        /* call f2 */
@@ -694,8 +703,7 @@ f2:
        movw h0, r20
        movw r28, r22
        rcall load32_from_X
-       movw acc0, r22
-       movw acc2, r24
+       rcall mov32_to_acc
        ldi r17, 15
 10:    rcall load32_from_X
        rcall eor32_to_acc
@@ -723,10 +731,8 @@ f2:
        movw r26, q16_0
        ldi r17, 16
 10:
-       ld acc0, Y+
-       ld acc1, Y+
-       ld acc2, Y+
-       ld acc3, Y+
+       rcall load32_from_Y
+       rcall mov32_to_acc
 ;---
        movw r22, xh0
        movw r24, xh2
@@ -785,9 +791,7 @@ f2:
        movw acc2, xl2
        rcall load32_from_X
        rcall eor32_to_acc
-       rcall load32_from_Y
-       rcall eor32_to_acc
-       rcall add_acc_to_Z
+       rcall eor_acc_from_Y_add_to_Z
        dec r17
        brne 10b
        sbiw r26, 9*4 /* X points to q[23] */
@@ -795,9 +799,7 @@ f2:
        eor acc1, xl0
        eor acc2, xl1
        eor acc3, xl2
-       rcall load32_from_Y
-       rcall eor32_to_acc
-       rcall add_acc_to_Z
+       rcall eor_acc_from_Y_add_to_Z
 ;---
        sbiw r26, 8*4 /* X points to q[16] */
        mov h0, r30
@@ -817,11 +819,9 @@ f2:
        rjmp 21f
 20:    rcall shiftright32
 21:
-       rcall eor32_to_acc
-       rcall load32_from_Y
-       rcall eor32_to_acc
        movw r30, h0
-       rcall add_acc_to_Z
+       rcall eor32_to_acc
+       rcall eor_acc_from_Y_add_to_Z
        movw h0, r30
        dec r17
        brne 10b
@@ -835,8 +835,7 @@ f2:
        rcall load32_from_X
        mov r20, r18
        rcall rotateleft32
-       movw acc0, r22
-       movw acc2, r24
+       rcall mov32_to_acc
        rcall add_acc_to_Z
        inc r18
        cpi r17, 5
@@ -997,16 +996,13 @@ bmw256_lastBlock:
        adc r23, r1
        adc r24, r1
        adc r25, r1
-       movw r30, buf0
-       adiw r30, 64-8
-       st Z+, r20
-       st Z+, r21
-       st Z+, r22
-       st Z+, r23
-       st Z+, r24
-       st Z+, r25
-       st Z+, r1
-       st Z+, r1
+       movw r26, buf0
+       adiw r26, 64-8
+       st X+, r20
+       st X+, r21
+       rcall store32_to_X
+       st X+, r1
+       st X+, r1
        movw r24, ctx0
        movw r22, buf0
        rcall bmw_small_nextBlock