the10thwiz/cc65
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Conio: 65c02 optimisations

Browse Source 
cputc: -2 cycles per char, -2 cycles per carriage return
cputs: -5 cycles per char, -3 cycles on cputsxy
vcprintf: -6 cycles per char
This commit is contained in:
Colin Leroy-Mira
2023-09-02 11:08:15 +02:00
parent d70c902a02
commit 8a140e6503
3 changed files with 69 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
libsrc/conio/cputs.s
View File

@@ -8,28 +8,46 @@
.export _cputsxy, _cputs
.import gotoxy, _cputc
.importzp ptr1, tmp1
.macpack cpu
_cputsxy:
sta ptr1 ; Save s for later
stx ptr1+1
jsr gotoxy ; Set cursor, pop x and y
.if (.cpu .bitand CPU_ISET_65SC02)
bra L0 ; Same as cputs...
.else
jmp L0 ; Same as cputs...
.endif
_cputs: sta ptr1 ; Save s
stx ptr1+1
L0: ldy #0
L1: lda (ptr1),y
beq L9 ; Jump if done
.if (.cpu .bitand CPU_ISET_65SC02)
L0: lda (ptr1) ; (5)
beq L9 ; (7) Jump if done
jsr _cputc ; (13) Output char, advance cursor
inc ptr1 ; (18) Bump low byte
bne L0 ; (20) Next char
inc ptr1+1 ; (25) Bump high byte
bne L0
.else
L0: ldy #0 ; (2)
L1: lda (ptr1),y ; (7)
beq L9 ; (9) Jump if done
iny
sty tmp1 ; Save offset
jsr _cputc ; Output char, advance cursor
ldy tmp1 ; Get offset
bne L1 ; Next char
inc ptr1+1 ; Bump high byte
sty tmp1 ; (14) Save offset
jsr _cputc ; (20) Output char, advance cursor
ldy tmp1 ; (23) Get offset
bne L1 ; (25) Next char
inc ptr1+1 ; (30) Bump high byte
bne L1
.endif
; Done
L9: rts

37
libsrc/conio/vcprintf.s
View File

@@ -10,7 +10,7 @@
.importzp sp, ptr1, ptr2, ptr3, tmp1
.macpack generic
.macpack cpu
.data
@@ -74,21 +74,40 @@ out: jsr popax ; count
; Loop outputting characters
.if (.cpu .bitand CPU_ISET_65SC02)
@L1: dec outdesc+6
beq @L4
@L2: ldy tmp1
lda (ptr1),y
iny
bne @L3
inc ptr1+1
@L3: sty tmp1
jsr _cputc
jmp @L1
@L2: lda (ptr1) ; (5)
inc ptr1 ; (10)
bne @L3 ; (12)
inc ptr1+1 ; (17)
@L3: jsr _cputc ; (23)
bra @L1 ; (26)
@L4: dec outdesc+7
bne @L2
rts
.else
@L1: dec outdesc+6
beq @L4
@L2: ldy tmp1 ; (3)
lda (ptr1),y ; (8)
iny ; (10)
bne @L3 ; (12)
inc ptr1+1 ; (17)
@L3: sty tmp1 ; (20)
jsr _cputc ; (26)
jmp @L1 ; (32)
@L4: dec outdesc+7
bne @L2
rts
.endif
; ----------------------------------------------------------------------------
; vcprintf - formatted console i/o
;