further shrinked bmw tiny to 1750 bytes

diff --git a/bmw/bmw_small-tinyasm.S b/bmw/bmw_small-tinyasm.S
index fca301a39e093fa83d89ba06059d4e1697c55352..7cdbadb59c4ebe93093a5bd09e594aee1c0405a2 100644 (file)
--- a/bmw/bmw_small-tinyasm.S
+++ b/bmw/bmw_small-tinyasm.S
@@ -338,8 +338,8 @@ load_rotate_add_M:
        inc r20
        rcall rotateleft32
        brts 10f
-       rcall add32_to_acc
-       ret
+       rjmp add32_to_acc
+;      ret
 10:    sub acc0, r22
        sbc acc1, r23
        sbc acc2, r24
@@ -386,8 +386,8 @@ addelement:
 load_sn_add:
        rcall load32_from_X
        rcall sn
-       rcall add32_to_acc
-       ret
+       rjmp add32_to_acc
+;      ret
 
 /*
   param q: r26:r27
@@ -397,11 +397,12 @@ load_sn_add:
 */
 
 expand_intro:
-       push_range 20, 27
-;      push r24
+
+       push_range 26, 27
+       push r24
        rcall addelement
-;      pop r24
-       pop_range 20, 27
+       pop r24
+       pop_range 26, 27
        lsl r24
        lsl r24
        add r26, r24
@@ -457,8 +458,8 @@ expand2:
        rcall load_sn_add
 expand2_exit:
        adiw r26, 4
-       rcall store_acc_to_dec_X
-       ret
+       rjmp store_acc_to_dec_X
+;      ret
 
 /******************************************************************************/
 /*
@@ -932,15 +933,14 @@ bmw256_lastBlock:
        memcpy(pctx.buffer, block, (length_b+7)/8);
        pctx.buffer[length_b>>3] |= 0x80 >> (length_b&0x07);
 */     movw r24, len0
+       ldi r23, 63
+       movw r26, blc0
        lsr r25
        ror r24
        lsr r24
        lsr r24
-       ldi r23, 63
-       sub r23, r24
-       movw r26, blc0
-       tst r24
        breq 301f
+       sub r23, r24
        /* copy (#r24) bytes to stack buffer */
 30: ld r20, X+
        st Z+, r20