Improve allocation of VRAM/PAL/OAM buffers (#222)

* Improve allocation of VRAM/PAL/OAM buffers Fixes #215: - Use weak symbols - Move PAL_BUF out of hardware stack - Split multi_vram* ops from the definition of the buffer * Fix weakness of VRAM_BUF
llvm-mos · Sep 30, 2023 · e44cf72 · e44cf72
1 parent 26b23c9
commit e44cf72
Show file tree

Hide file tree

Showing 8 changed files with 112 additions and 118 deletions.
diff --git a/mos-platform/nes/nesdoug/CMakeLists.txt b/mos-platform/nes/nesdoug/CMakeLists.txt
@@ -12,6 +12,7 @@ add_platform_library(nes-nesdoug
   padlib.s
   vram_buffer.c
   vram_buffer.s
+  vram_buffer_ops.s
   zaplib.s
 )
 target_include_directories(nes-nesdoug BEFORE PUBLIC ..)

diff --git a/mos-platform/nes/nesdoug/vram_buffer.c b/mos-platform/nes/nesdoug/vram_buffer.c
@@ -3,11 +3,12 @@
 
 __attribute__((section(".zp.vram_index"))) char VRAM_INDEX;
 
-extern char VRAM_BUF[];
+extern char VRAM_BUF[128];
+
+__attribute__((leaf)) void __post_vram_update();
 
 void set_vram_update(const void *buf);
 void set_vram_buffer(void) {
-  VRAM_BUF[0] = 0xff;
-  VRAM_INDEX = 0;
   set_vram_update(VRAM_BUF);
+  __post_vram_update();
 }
diff --git a/mos-platform/nes/nesdoug/vram_buffer.s b/mos-platform/nes/nesdoug/vram_buffer.s
@@ -4,10 +4,10 @@
 .zeropage VRAM_INDEX
 
 .section .noinit.vram_buf,"aw",@nobits
-.globl VRAM_BUF
+.weak VRAM_BUF
 .balign 128
 VRAM_BUF:
-  .zero 128
+	.zero 128
 
 .text
 .global __post_vram_update
@@ -17,102 +17,3 @@ __post_vram_update:
 	inx ;x=0
 	stx VRAM_INDEX
 	rts
-
-;void multi_vram_buffer_horz(char * data, char len, int ppu_address);
-.section .text.multi_vram_buffer_horz,"ax",@progbits
-.globl multi_vram_buffer_horz
-multi_vram_buffer_horz:
-	;     A - len
-	;     X - <ppu_address
-	; __rc2 - <data
-	; __rc3 - >data
-	; __rc4 - >ppu_address
-	ldy VRAM_INDEX
-
-	sta __rc5         ; save len for loop comparison
-	sta VRAM_BUF+2, y ; store len in header
-
-	lda #$40          ; load horizontal flag
-
-multi_vram_buffer_common:
-	ora __rc4         ; combine direction flag with upper address byte
-	sta VRAM_BUF+0, y ; store upper byte in header
-	txa
-	sta VRAM_BUF+1, y ; store lower byte in header
-
-	; get data index past header
-	tya
-	clc
-	adc #3
-	tax ;need y for source, x is for dest and for vram_index
-
-	ldy #0
-
-  ; TODO: unroll?
-.Lmulti_vram_buffer_common_loop:
-	lda (__rc2), y
-	sta VRAM_BUF, x
-	inx
-	iny
-	cpy __rc5
-	bne .Lmulti_vram_buffer_common_loop
-	lda #$ff ;=NT_UPD_EOF
-	sta VRAM_BUF, x
-	stx VRAM_INDEX
-	rts
-
-
-
-
-;void multi_vram_buffer_vert(char * data, char len, int ppu_address);
-.section .text.multi_vram_buffer_vert,"ax",@progbits
-.globl multi_vram_buffer_vert
-multi_vram_buffer_vert:
-	;     A - len
-	;     X - <ppu_address
-	; __rc2 - <data
-	; __rc3 - >data
-	; __rc4 - >ppu_address
-	ldy VRAM_INDEX
-
-	sta __rc5         ; save len for loop comparison
-	sta VRAM_BUF+2, y ; store len in header
-
-	lda #$80          ; load vertical flag
-	bne multi_vram_buffer_common ; always taken
-
-
-
-
-;void one_vram_buffer(char data, int ppu_address);
-.section .text.one_vram_buffer,"ax",@progbits
-.globl one_vram_buffer
-one_vram_buffer:
-	;     A - data
-	;     X - <ppu_address
-	; __rc2 - >ppu_address
-	ldy VRAM_INDEX
-
-	sta VRAM_BUF+2, y ; data
-	lda __rc2
-	sta VRAM_BUF, y   ; address hi
-	txa
-	sta VRAM_BUF+1, y ; adderess lo
-	iny
-	iny
-	iny
-	lda #$ff ;=NT_UPD_EOF
-	sta VRAM_BUF, y
-	sty VRAM_INDEX
-	rts
-
-
-
-
-;void clear_vram_buffer(void);
-;_clear_vram_buffer:
-;	lda #0
-;	sta VRAM_INDEX
-;	lda #$ff
-;	sta VRAM_BUF
-;	rts
diff --git a/mos-platform/nes/nesdoug/vram_buffer_ops.s b/mos-platform/nes/nesdoug/vram_buffer_ops.s
@@ -0,0 +1,103 @@
+;written by Doug Fraker
+;version 1.2, 1/1/2022
+
+.zeropage VRAM_INDEX
+
+; Horizontal and vertical share common code and need to stay in the same section
+.section .text.multi_vram_buffer,"ax",@progbits
+
+;void multi_vram_buffer_horz(char * data, char len, int ppu_address);
+.globl multi_vram_buffer_horz
+multi_vram_buffer_horz:
+	;     A - len
+	;     X - <ppu_address
+	; __rc2 - <data
+	; __rc3 - >data
+	; __rc4 - >ppu_address
+	ldy VRAM_INDEX
+
+	sta __rc5         ; save len for loop comparison
+	sta VRAM_BUF+2, y ; store len in header
+
+	lda #$40          ; load horizontal flag
+
+multi_vram_buffer_common:
+	ora __rc4         ; combine direction flag with upper address byte
+	sta VRAM_BUF+0, y ; store upper byte in header
+	txa
+	sta VRAM_BUF+1, y ; store lower byte in header
+
+	; get data index past header
+	tya
+	clc
+	adc #3
+	tax ;need y for source, x is for dest and for vram_index
+
+	ldy #0
+
+  ; TODO: unroll?
+.Lmulti_vram_buffer_common_loop:
+	lda (__rc2), y
+	sta VRAM_BUF, x
+	inx
+	iny
+	cpy __rc5
+	bne .Lmulti_vram_buffer_common_loop
+	lda #$ff ;=NT_UPD_EOF
+	sta VRAM_BUF, x
+	stx VRAM_INDEX
+	rts
+
+
+
+;void multi_vram_buffer_vert(char * data, char len, int ppu_address);
+.globl multi_vram_buffer_vert
+multi_vram_buffer_vert:
+	;     A - len
+	;     X - <ppu_address
+	; __rc2 - <data
+	; __rc3 - >data
+	; __rc4 - >ppu_address
+	ldy VRAM_INDEX
+
+	sta __rc5         ; save len for loop comparison
+	sta VRAM_BUF+2, y ; store len in header
+
+	lda #$80          ; load vertical flag
+	bne multi_vram_buffer_common ; always taken
+
+
+
+
+;void one_vram_buffer(char data, int ppu_address);
+.section .text.one_vram_buffer,"ax",@progbits
+.globl one_vram_buffer
+one_vram_buffer:
+	;     A - data
+	;     X - <ppu_address
+	; __rc2 - >ppu_address
+	ldy VRAM_INDEX
+
+	sta VRAM_BUF+2, y ; data
+	lda __rc2
+	sta VRAM_BUF, y   ; address hi
+	txa
+	sta VRAM_BUF+1, y ; adderess lo
+	iny
+	iny
+	iny
+	lda #$ff ;=NT_UPD_EOF
+	sta VRAM_BUF, y
+	sty VRAM_INDEX
+	rts
+
+
+
+; now in __post_vram_update
+;void clear_vram_buffer(void);
+;_clear_vram_buffer:
+;	lda #0
+;	sta VRAM_INDEX
+;	lda #$ff
+;	sta VRAM_BUF
+;	rts
diff --git a/mos-platform/nes/neslib/oam_update.c b/mos-platform/nes/neslib/oam_update.c
@@ -4,6 +4,7 @@
 // information.
 
 __attribute__((section(".zp.sprid"))) unsigned SPRID;
+__attribute__((weak, aligned(256), section(".noinit.oam_buf"))) char OAM_BUF[256];
 
 void oam_set(char index) { SPRID = index & 0xfc; }
 

diff --git a/mos-platform/nes/neslib/oam_update.s b/mos-platform/nes/neslib/oam_update.s
@@ -7,14 +7,6 @@
 
 .include "nes.inc"
 
-; Reserve space at beginning of RAM for OAM buffer.
-.section .noinit.oam_buf,"a",@nobits
-.globl OAM_BUF
-.align 256
-OAM_BUF:
-  .space 256
-
-
 .section .init.270,"axR",@progbits
 .globl oam_update_init
 oam_update_init:

diff --git a/mos-platform/nes/neslib/pal_update.c b/mos-platform/nes/neslib/pal_update.c
@@ -4,9 +4,8 @@
 // information.
 
 __attribute__((section(".zp.pal_update"))) volatile char PAL_UPDATE;
+__attribute__((weak, aligned(32), section(".noinit.pal_buf"))) volatile char PAL_BUF[32];
 
-
-extern volatile char PAL_BUF[];
 void pal_col(char index, char color) {
   PAL_BUF[index & 0x1f] = color;
   ++PAL_UPDATE;

diff --git a/mos-platform/nes/neslib/pal_update.s b/mos-platform/nes/neslib/pal_update.s
@@ -7,10 +7,6 @@
 
 .include "nes.inc"
 
-; Place the pallette buffer at the very bottom 32 bytes of the hard stack.
-.globl PAL_BUF
-PAL_BUF = 0x0100
-
 .section .init.265,"axR",@progbits
 .globl pal_update_init
 pal_update_init: