From 2e8dfde2904e5cc7ec5e4827c030749c5bb48502 Mon Sep 17 00:00:00 2001
From: blondie7575 <qd@quinndunki.com>
Date: Sat, 29 Jul 2017 14:47:17 -0700
Subject: [PATCH] Fast graphics path screen fill working

---
 Makefile    |   2 +-
 V2Make.scpt | Bin 2462 -> 2462 bytes
 equates.s   |  14 ++++---
 graphics.s  |  45 +++++++++++++++++------
 gscats.2mg  | Bin 819264 -> 819264 bytes
 gscats.s    |  45 +++++++++++++++++++++--
 macros.s    | 104 +++++++++++++++++++++++++++++++++++++++++++---------
 7 files changed, 171 insertions(+), 39 deletions(-)

diff --git a/Makefile b/Makefile
index b6a348d..4e2e63f 100644
--- a/Makefile
+++ b/Makefile
@@ -11,7 +11,7 @@
 
 CL65=cl65
 AC=AppleCommander.jar
-ADDR=6000
+ADDR=800
 
 PGM=gscats
 
diff --git a/V2Make.scpt b/V2Make.scpt
index 274dd79d26988fcdfc490c50cd3b8ef30063a3b2..cdcc6ea32183ceca9a0ec53f23a2d33f0f2fddf1 100644
GIT binary patch
delta 22
ecmbOyJWqH-J{!}$?~@DJ{Fss=HXmfW%LD*rz6i_!

delta 22
ecmbOyJWqH-J{!|-@yUg3eoS>0n-8+xWdZ<QcnC!R

diff --git a/equates.s b/equates.s
index 4c06ca1..437135d 100644
--- a/equates.s
+++ b/equates.s
@@ -1,10 +1,12 @@
-TEXTCOLOR		= $c022
-NEWVIDEO		= $c029
-BORDERCOLOR		= $c034
-CLOCKCTL		= $c034
+TEXTCOLOR		= $e0c022
+NEWVIDEO		= $e0c029
+BORDERCOLOR		= $e0c034
+CLOCKCTL		= $e0c034
+SHADOW			= $e0c035
+STACKCTL		= $e0c068
+KBD				= $e0c000
+KBDSTROBE		= $e0c010
 COUT			= $fded
-KBD				= $c000
-KBDSTROBE		= $c010
 
 
 ; Zero page locations we use (unused by Monitor, Applesoft, or ProDOS)
diff --git a/graphics.s b/graphics.s
index fa80fc7..ae20319 100644
--- a/graphics.s
+++ b/graphics.s
@@ -1,19 +1,31 @@
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; colorFill
-; Fills the screen with a color (or two). Not very fast.
+; Fills the screen with a color (or two). Pretty fast, but not fastest possible
 ; A 4:4:4:4 = Palette entries
+;
+; Trashes Y
 
 colorFill:
-	phx
-	ldx #32000
+	FASTGRAPHICS
+
+	lda #$9d00-1	; Point stack to end of VRAM
+	tcs
+
+	ldy #200
 
 colorFillLoop:
-	sta $e11ffe,x
-	dex
-	dex
-	bne colorFillLoop
-	plx
+	; 80 PHXs, for 1 line
+	; We could do the entire screen with PHXs, but this is a
+	; balance between speed and super-verbose code
+	.byte $da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da
+	.byte $da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da
+	.byte $da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da
+	.byte $da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da,$da
 
+	dey
+	bne colorFillLoop
+
+	SLOWGRAPHICS
 	rts
 
 
@@ -30,7 +42,7 @@ initSCBs:
 initSCBsLoop:
 	dex
 	dex
-	sta $e19cfe,x
+	sta $e19d00,x
 	bne initSCBsLoop
 	rts
 
@@ -87,8 +99,19 @@ setPaletteLoop:
 	iny
 	inx
 	inx
-	cmp #16
+	cpx #32
 	bne setPaletteLoop
 
 	RESTORE_XY
 	rts
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; caches
+shadowRegister:
+	.byte 0
+stackRegister:
+	.byte 0
+stackPtr:
+	.word 0
+
diff --git a/gscats.2mg b/gscats.2mg
index 8809fe8ce34d165510ea1870cd854d9fb867653f..df51020b5e604152b32150654782e4a13333f4ce 100644
GIT binary patch
delta 404
zcmb`?ze~eV5C`x}?2mg2Hra#@zIN*p(o&(Ipw7kJF^eFAtBBC1qM<G!l&(spi#$q*
z;NT=Yqk?qmEFB)194l@uqF4Gac<?#Ck2^S?Y*l2da(|RI&o%fdTB`srfb*btclEc2
zolL6f_5nl{eG{Lm*aOr*)t)x`7A5q{I3HLzwy-{^%^Vs6Kz)$dViFXkifz8d5<4nM
z#hwG|u`Sx@Wds0sSl=}H9r6=gY?QVa8{;z8W)VnbOanuzH8zWF2XYh0bqaa+G~v;P
zRDOdA2q;8{^k_0>2NYty>AOt)UgvM$;p3m{Y7{2C5H{HSY}jOrgQ3gv`wo4m5880p
sgPZQ_rIl42Mqg*sZs&OY!tmOLrxBejkrYXj49OBCj4VGI-paxB4}#IdsQ>@~

delta 189
zcmX@mV056tXhRmO;4TJ+KS~^e3=D@E5<m>5$&O4%H#@Mg)cLIZz|gC7K%p|>k-=Kc
z1DpqYH4hvzSm~Y6%C?e$L1AvfN@HV%>;wfHh`gpfSl;4yg2nH(3<m^$_X-@4_<hJA
zfr0-61OG%t>B$?KWG8pGYO$XbV7TzH>eu9nx$?~yTH7zQG6FFZ5HkZY3lOscF&hxG
X12G2>a{@8f_6x1t1-~W-Fzo;Uj(<;v

diff --git a/gscats.s b/gscats.s
index 7ff91bc..aff6e36 100644
--- a/gscats.s
+++ b/gscats.s
@@ -10,13 +10,49 @@
 .include "equates.s"
 
 
-.org $6000
+.org $800
 
 main:
 	NATIVE
 
+mainCopyStart:
+	ldx #0
+	lda #mainBank2
+	sta mainCopyDest+1
+
+mainCopyLoop:
+	lda mainBank2,x
+
+mainCopyDest:
+	sta $020800,x
+	inx
+	cpx #endMainBank2-mainBank2
+	bne mainCopyLoop
+
+	lda #returnToProDOS
+	sta proDOSLongJump
+	lda #mainBank2
+	sta mainLongJump
+	jml (mainLongJump)
+
+returnToProDOS:
+	SYNCDBR
+	EMULATION
+	rts
+
+mainLongJump:
+	.byte 00,08,02
+proDOSLongJump:
+	.byte 00,00,00
+
+
+mainBank2:
+	SYNCDBR
+
+	BITS8
 	lda #$f0
 	sta TEXTCOLOR
+	BITS16
 
 	jsr initSCBs
 	SHRVIDEO
@@ -26,14 +62,13 @@ main:
 	lda #0
 	jsr setPalette
 	
-	lda #$3333
+	ldx #$1111
 	jsr colorFill
 
 	jsr kbdWait
 	CLASSICVIDEO
 
-	EMULATION
-	rts
+	jml (proDOSLongJump)
 
 
 
@@ -62,6 +97,8 @@ basePalette:
 .include "graphics.s"
 
 
+endMainBank2:
+
 
 ; Suppress some linker warnings - Must be the last thing in the file
 ; This is because Quinn doesn't really know how to use ca65 properly
diff --git a/macros.s b/macros.s
index 67c1dee..31a1f58 100644
--- a/macros.s
+++ b/macros.s
@@ -1,44 +1,114 @@
+.macro OP8
+	.i8
+	.a8
+.endmacro
+
+
+.macro OP16
+	.i16
+	.a16
+.endmacro
+
+
+.macro BITS8
+	sep #%00110000
+	OP8
+.endmacro
+
+
+.macro BITS16
+	rep #%00110000
+	OP16
+.endmacro
+
+
+.macro DBR bankNum
+	BITS8
+	lda bankNum
+	pha
+	plb
+	BITS16
+.endmacro
+
+
+.macro SYNCDBR
+	phk
+	plb
+.endmacro
+
+
 .macro EMULATION
 	sec		; Enable 8-bit mode
 	xce
-	.i8
-	.a8
+	OP8
 .endmacro
 
 
 .macro NATIVE
 	clc				; Enable 16-bit mode
 	xce
-	rep	#$30
-	.i16
-	.a16
+	BITS16
 .endmacro
 
+
 .macro SHRVIDEO
-	sep #$30
-	.i8
-	.a8
+	BITS8
 	lda NEWVIDEO
 	ora	#%11000001
 	sta NEWVIDEO
-	rep	#$30
-	.i16
-	.a16
+	BITS16
 .endmacro
 
+
 .macro CLASSICVIDEO
-	sep #$30
-	.i8
-	.a8
+	BITS8
 	lda NEWVIDEO
 	and #%00111111
 	sta NEWVIDEO
-	rep	#$30
-	.i16
-	.a16
+	BITS16
 .endmacro
 
 
+.macro FASTGRAPHICS
+	sei
+	sep #%00100000	; 16-bit A only, to preserve X/Y
+	.a8
+
+	lda SHADOW
+	sta shadowRegister
+	lda #0
+	sta SHADOW
+
+	lda STACKCTL
+	sta stackRegister
+	ora #$30
+	sta STACKCTL
+
+	rep #%00100000
+	.a16
+	tsc
+	sta stackPtr
+.endmacro
+
+
+.macro SLOWGRAPHICS
+	sep #%00100000	; 16-bit A only, to preserve X/Y
+	.a8
+
+	lda shadowRegister
+	sta SHADOW
+
+	lda stackRegister
+	sta STACKCTL
+
+	rep #%00100000
+	.a16
+	lda stackPtr
+	tcs
+
+	cli
+.endmacro
+
 
 
 ;;;;;;;;;;