further shrinked bmw tiny to 1648 bytes

diff --git a/bmw/bmw_small-tinyasm.S b/bmw/bmw_small-tinyasm.S
index 94f1825a7b50704351877249f826f7aeff9ac760..50e110d82d9635c14a735ee176cc30d6e919b92f 100644 (file)
--- a/bmw/bmw_small-tinyasm.S
+++ b/bmw/bmw_small-tinyasm.S
@@ -127,14 +127,7 @@ s2:  .byte 2, 1,12,25
 s3:  .byte 2, 2,15,29
 s4:  .byte 1, 0, 0, 0
 s5:  .byte 2, 0, 0, 0
-/*
-s0:  .byte 0x34, 19
-s1:  .byte 0x28, 23
-s2:  .byte 0x9C, 25
-s3:  .byte 0xAF, 29
-s4:  .byte 0x00,  0
-s5:  .byte 0x80,  0
-*/
+
 h0   = 10
 h1   = 11
 m0   = 12
@@ -177,8 +170,7 @@ sn:
        pop acc2
        pop acc1
        pop acc0
-       pop_range 2, 5
-       ret
+       rjmp pop5
 
 /******************************************************************************/
 /*
@@ -506,21 +498,28 @@ h0 = 4
 h1 = 5
 m0 = 6
 m1 = 7
+ctx0 =   2
+ctx1 =   3
+msg0 =   4
+msg1 =   5
 
 restore_f1:
        movw r26, r2
        movw r22, r4
     movw r20, r6
        ret
-
+bmw_small_nextBlock_early:
+       movw r24, ctx0
+       movw r22, msg0
 .global bmw_small_nextBlock
 .global bmw224_nextBlock
 .global bmw256_nextBlock
 bmw_small_nextBlock:
 bmw224_nextBlock:
 bmw256_nextBlock:
+       push_range  2, 7
        push_range 28, 29
-       push_range  2, 17
+       push_range  8, 17
        stack_alloc_large 32*4, r28, r29
        ldi r16, 0x4f
        push r16
@@ -851,8 +850,15 @@ f2:
 ;      rcall printX
 ;--- END DBG
        stack_free_large3 32*4+4
-       pop_range  2, 17
+       pop_range 10, 17
+pop9:
+       pop_range 8, 9
+pop28:
        pop_range 28, 29
+pop7:
+       pop_range 6, 7
+pop5:
+       pop_range 2, 5
        ret
 
 /******************************************************************************/
@@ -901,9 +907,7 @@ bmw256_lastBlock:
 1:
        cpi len1, hi8(512)
        brlo 2f
-       movw r24, ctx0
-       movw r22, blc0
-       rcall bmw_small_nextBlock
+       rcall bmw_small_nextBlock_early
        ldi r24, 64
        add blc0, r24
        adc blc1, r1
@@ -965,9 +969,8 @@ bmw256_lastBlock:
        breq 400f
        cpi len0, 192
        brlo 400f
-       movw r24, ctx0
-       movw r22, buf0
-       rcall bmw_small_nextBlock
+       movw blc0, buf0
+       rcall bmw_small_nextBlock_early
        movw r26, buf0
        ldi r20, 64-8
 350:
@@ -1003,9 +1006,8 @@ bmw256_lastBlock:
        rcall store32_to_X
        st X+, r1
        st X+, r1
-       movw r24, ctx0
-       movw r22, buf0
-       rcall bmw_small_nextBlock
+       movw blc0, buf0
+       rcall bmw_small_nextBlock_early
 /*     memset(pctx.buffer, 0xaa, 64);
        for(i=0; i<16;++i){
                pctx.buffer[i*4] = i+0xa0;
@@ -1024,9 +1026,9 @@ bmw256_lastBlock:
 /*     bmw_small_nextBlock((bmw_small_ctx_t*)&pctx, ctx->h);
        memcpy(ctx->h, pctx.buffer, 64);
 */
-       movw r24, buf0
-       movw r22, ctx0
-       rcall bmw_small_nextBlock
+    movw r24, buf0
+    movw r22, ctx0
+    rcall bmw_small_nextBlock
        ldi r18, 64
        movw r26, ctx0
        movw r30, buf0
@@ -1037,9 +1039,7 @@ bmw256_lastBlock:
        brne 600b
 
        stack_free_large 68
-       pop_range 28, 29
-       pop_range 2, 7
-       ret
+       rjmp pop28
 
 
 /*******************************************************************************
@@ -1106,10 +1106,10 @@ dst0 =  10
 dst1 =  11
 .global bmw256
 bmw256:
-       push r16
-       ldi r16, 1
+       set
        rjmp bmw_small_all
 
+
 /*******************************************************************************
 * void bmw224(void* dest, const void* msg, uint32_t length_b){
 *      bmw_small_ctx_t ctx;
@@ -1139,15 +1139,21 @@ dst0 =   6
 dst1 =   7
 .global bmw224
 bmw224:
-       push r16
-       clr r16
+       clt
+       rjmp bmw_small_all
+
 
 bmw_small_all:
-       push_range 2, 9
+       push_range 2, 7
        push_range 28, 29
+       push_range 8, 9
+       push r16
        stack_alloc_large 64+4
        adiw r30, 1
-       movw ctx0, r30
+       clr r16
+       brtc 10f
+       inc r16
+10:    movw ctx0, r30
        movw dst0, r24
        movw msg0, r22
        movw len0, r18
@@ -1162,9 +1168,7 @@ bmw_small_all:
        mov r18, len2
        or  r18, len3
        breq 50f
-       movw r24, ctx0
-       movw r22, msg0
-       rcall bmw_small_nextBlock
+       rcall bmw_small_nextBlock_early
        subi len1, 2
        sbc len2, r1
        sbc len3, r1
@@ -1185,10 +1189,8 @@ bmw_small_all:
        adc r31, r1
        icall
        stack_free_large 64+4
-       pop_range 28, 29
-       pop_range 2, 9
        pop r16
-       ret
+       rjmp pop9
 
 init_lut:
        rjmp bmw224_init