improved delay() only needs t0 running

Author: nerdralph

src/delay.S

@@ -6,42 +6,23 @@
 #include <avr/io.h>
 #include "macros.inc"
 
-;struct {
-;    uint32_t t0_ovfl;
-;    uint8_t ovfl_pad;
-;} t0_millis;
-
-
-; to force t0_isr to get linked in
-.global t0_ovfl_isr
-
-; read lower 16 bits of extended TCNT0 into r19:r18
-r18_t0cnt:
-    in r18, TCNT0
-    lds r19, t0_millis
-    in r0, TCNT0
-    sub r0, r18                         ; TCNT0 overflow?
-    brcs r18_t0cnt
-ret
-
-; delay for r27:r24 (ulong) ms
+#define remain  r19
+#define start   r20
+#define current r21
+; delay for r26:r24 (u24) ms
 GLABEL delay_impl
-    rcall r18_t0cnt
-    movw r20, r18                       ; r20 = start count
-    rjmp decms
-dloop:
-    rcall r18_t0cnt
-    sub r18, r20                        ; current - start
-    sbc r19, r21
-    subi r18, lo8(T0CNT_PER_MS)         ; > 1ms?
-    sbci r19, hi8(T0CNT_PER_MS)
-    brcc dloop
-    subi r20, lo8(-(T0CNT_PER_MS))        ; add 1ms to start
-    sbci r21, hi8(-(T0CNT_PER_MS))
+    in  start, TCNT0
+    ldi remain, lo8(T0CNT_PER_MS)
+wait1ms:                                ; check TCNT0 until 1ms has passed
+    in  current, TCNT0
+    sub current, start
+    add start, current                  ; advance start
+    sub remain, current
+    brcc wait1ms
+    subi remain, lo8(-(T0CNT_PER_MS))   ; add 1ms
 decms:                                  ; decrement ms argument
     sbiw r24, 1
     sbci r26, 0
-    sbci r27, 0
-    brcc dloop
+    brne wait1ms
     ret