Port blit from Emerald

8 files changed, 233 insertions, 830 deletions

diff --git a/asm/blit.s b/asm/blit.s
deleted file mode 100644
index 84b56fd67..000000000
--- a/asm/blit.s
+++ /dev/null
@@ -1,813 +0,0 @@
-	.include "asm/macros.inc"
-	.include "constants/constants.inc"
-
-	.syntax unified
-
-	.text
-
-	thumb_func_start BlitBitmapRect4BitWithoutColorKey
-BlitBitmapRect4BitWithoutColorKey: @ 8004A60
-	push {r4-r7,lr}
-	mov r7, r8
-	push {r7}
-	sub sp, 0x14
-	ldr r4, [sp, 0x2C]
-	ldr r5, [sp, 0x30]
-	ldr r6, [sp, 0x34]
-	ldr r7, [sp, 0x38]
-	lsls r2, 16
-	lsrs r2, 16
-	lsls r3, 16
-	lsrs r3, 16
-	lsls r4, 16
-	lsrs r4, 16
-	lsls r5, 16
-	lsrs r5, 16
-	lsls r6, 16
-	lsrs r6, 16
-	lsls r7, 16
-	lsrs r7, 16
-	str r4, [sp]
-	str r5, [sp, 0x4]
-	str r6, [sp, 0x8]
-	str r7, [sp, 0xC]
-	movs r4, 0xFF
-	str r4, [sp, 0x10]
-	bl BlitBitmapRect4Bit
-	add sp, 0x14
-	pop {r3}
-	mov r8, r3
-	pop {r4-r7}
-	pop {r0}
-	bx r0
-	thumb_func_end BlitBitmapRect4BitWithoutColorKey
-
-	thumb_func_start BlitBitmapRect4Bit
-BlitBitmapRect4Bit: @ 8004AA4
-	push {r4-r7,lr}
-	mov r7, r10
-	mov r6, r9
-	mov r5, r8
-	push {r5-r7}
-	sub sp, 0x2C
-	str r0, [sp]
-	str r1, [sp, 0x4]
-	ldr r0, [sp, 0x4C]
-	ldr r1, [sp, 0x50]
-	ldr r4, [sp, 0x54]
-	ldr r5, [sp, 0x58]
-	ldr r6, [sp, 0x5C]
-	lsls r2, 16
-	lsrs r2, 16
-	str r2, [sp, 0x8]
-	lsls r3, 16
-	lsrs r7, r3, 16
-	lsls r0, 16
-	lsrs r0, 16
-	str r0, [sp, 0xC]
-	lsls r1, 16
-	lsrs r1, 16
-	mov r9, r1
-	lsls r4, 16
-	lsrs r4, 16
-	lsls r5, 16
-	lsrs r5, 16
-	lsls r6, 24
-	lsrs r6, 24
-	str r6, [sp, 0x10]
-	ldr r0, [sp, 0x4]
-	ldrh r3, [r0, 0x4]
-	ldr r1, [sp, 0xC]
-	subs r0, r3, r1
-	adds r2, r4
-	mov r8, r2
-	cmp r0, r4
-	bge _08004AF8
-	ldr r2, [sp, 0x8]
-	adds r0, r2
-	mov r8, r0
-_08004AF8:
-	ldr r4, [sp, 0x4]
-	ldrh r0, [r4, 0x6]
-	mov r1, r9
-	subs r0, r1
-	adds r2, r5, r7
-	str r2, [sp, 0x14]
-	cmp r0, r5
-	bge _08004B0C
-	adds r0, r7
-	str r0, [sp, 0x14]
-_08004B0C:
-	ldr r4, [sp]
-	ldrh r1, [r4, 0x4]
-	movs r2, 0x7
-	adds r0, r1, 0
-	ands r0, r2
-	adds r1, r0
-	asrs r1, 3
-	str r1, [sp, 0x18]
-	adds r0, r3, 0
-	ands r0, r2
-	adds r0, r3, r0
-	asrs r0, 3
-	str r0, [sp, 0x1C]
-	ldr r0, [sp, 0x10]
-	cmp r0, 0xFF
-	bne _08004BCA
-	adds r3, r7, 0
-	mov r1, r9
-	ldr r2, [sp, 0x14]
-	cmp r3, r2
-	blt _08004B38
-	b _08004C72
-_08004B38:
-	ldr r5, [sp, 0x8]
-	ldr r6, [sp, 0xC]
-	adds r4, r3, 0x1
-	str r4, [sp, 0x24]
-	adds r0, r1, 0x1
-	str r0, [sp, 0x28]
-	cmp r5, r8
-	bge _08004BBE
-	movs r7, 0x3
-	asrs r0, r3, 3
-	ldr r2, [sp, 0x18]
-	muls r0, r2
-	lsls r0, 5
-	mov r10, r0
-	lsls r0, r3, 29
-	lsrs r0, 27
-	mov r9, r0
-	asrs r0, r1, 3
-	ldr r4, [sp, 0x1C]
-	muls r0, r4
-	lsls r0, 5
-	mov r12, r0
-	lsls r0, r1, 29
-	lsrs r3, r0, 27
-_08004B68:
-	asrs r0, r5, 1
-	ands r0, r7
-	ldr r2, [sp]
-	ldr r1, [r2]
-	adds r1, r0
-	asrs r0, r5, 3
-	lsls r0, 5
-	adds r1, r0
-	add r1, r10
-	mov r4, r9
-	adds r2, r1, r4
-	asrs r0, r6, 1
-	ands r0, r7
-	ldr r4, [sp, 0x4]
-	ldr r1, [r4]
-	adds r1, r0
-	asrs r0, r6, 3
-	lsls r0, 5
-	adds r1, r0
-	add r1, r12
-	adds r4, r1, r3
-	ldrb r2, [r2]
-	adds r0, r5, 0
-	movs r1, 0x1
-	ands r0, r1
-	lsls r0, 2
-	asrs r2, r0
-	movs r0, 0xF
-	ands r2, r0
-	adds r0, r6, 0
-	ands r0, r1
-	lsls r1, r0, 2
-	lsls r2, r1
-	movs r0, 0xF0
-	asrs r0, r1
-	ldrb r1, [r4]
-	ands r0, r1
-	orrs r2, r0
-	strb r2, [r4]
-	adds r5, 0x1
-	adds r6, 0x1
-	cmp r5, r8
-	blt _08004B68
-_08004BBE:
-	ldr r3, [sp, 0x24]
-	ldr r1, [sp, 0x28]
-	ldr r2, [sp, 0x14]
-	cmp r3, r2
-	blt _08004B38
-	b _08004C72
-_08004BCA:
-	adds r3, r7, 0
-	mov r1, r9
-	ldr r4, [sp, 0x14]
-	cmp r3, r4
-	bge _08004C72
-_08004BD4:
-	ldr r5, [sp, 0x8]
-	ldr r6, [sp, 0xC]
-	adds r0, r3, 0x1
-	str r0, [sp, 0x24]
-	adds r2, r1, 0x1
-	str r2, [sp, 0x28]
-	cmp r5, r8
-	bge _08004C68
-	movs r4, 0x3
-	mov r9, r4
-	asrs r0, r3, 3
-	ldr r2, [sp, 0x18]
-	muls r0, r2
-	lsls r0, 5
-	str r0, [sp, 0x20]
-	lsls r0, r3, 29
-	lsrs r0, 27
-	mov r10, r0
-	movs r3, 0x1
-	asrs r0, r1, 3
-	ldr r4, [sp, 0x1C]
-	muls r0, r4
-	lsls r0, 5
-	mov r12, r0
-	lsls r0, r1, 29
-	lsrs r7, r0, 27
-_08004C08:
-	asrs r0, r5, 1
-	mov r1, r9
-	ands r0, r1
-	ldr r2, [sp]
-	ldr r1, [r2]
-	adds r1, r0
-	asrs r0, r5, 3
-	lsls r0, 5
-	adds r1, r0
-	ldr r4, [sp, 0x20]
-	adds r1, r4
-	mov r0, r10
-	adds r2, r1, r0
-	asrs r0, r6, 1
-	mov r1, r9
-	ands r0, r1
-	ldr r4, [sp, 0x4]
-	ldr r1, [r4]
-	adds r1, r0
-	asrs r0, r6, 3
-	lsls r0, 5
-	adds r1, r0
-	add r1, r12
-	adds r4, r1, r7
-	ldrb r2, [r2]
-	adds r0, r5, 0
-	ands r0, r3
-	lsls r0, 2
-	asrs r2, r0
-	movs r0, 0xF
-	ands r2, r0
-	ldr r0, [sp, 0x10]
-	cmp r2, r0
-	beq _08004C60
-	adds r0, r6, 0
-	ands r0, r3
-	lsls r1, r0, 2
-	lsls r2, r1
-	movs r0, 0xF0
-	asrs r0, r1
-	ldrb r1, [r4]
-	ands r0, r1
-	orrs r2, r0
-	strb r2, [r4]
-_08004C60:
-	adds r5, 0x1
-	adds r6, 0x1
-	cmp r5, r8
-	blt _08004C08
-_08004C68:
-	ldr r3, [sp, 0x24]
-	ldr r1, [sp, 0x28]
-	ldr r2, [sp, 0x14]
-	cmp r3, r2
-	blt _08004BD4
-_08004C72:
-	add sp, 0x2C
-	pop {r3-r5}
-	mov r8, r3
-	mov r9, r4
-	mov r10, r5
-	pop {r4-r7}
-	pop {r0}
-	bx r0
-	thumb_func_end BlitBitmapRect4Bit
-
-	thumb_func_start FillBitmapRect4Bit
-FillBitmapRect4Bit: @ 8004C84
-	push {r4-r7,lr}
-	mov r7, r10
-	mov r6, r9
-	mov r5, r8
-	push {r5-r7}
-	sub sp, 0x4
-	mov r9, r0
-	ldr r0, [sp, 0x24]
-	ldr r4, [sp, 0x28]
-	lsls r1, 16
-	lsrs r1, 16
-	mov r10, r1
-	lsls r2, 16
-	lsrs r2, 16
-	adds r5, r2, 0
-	lsls r3, 16
-	lsrs r3, 16
-	lsls r0, 16
-	lsrs r0, 16
-	lsls r4, 24
-	lsrs r7, r4, 24
-	adds r4, r1, r3
-	mov r3, r9
-	ldrh r1, [r3, 0x4]
-	cmp r4, r1
-	ble _08004CBA
-	adds r4, r1, 0
-_08004CBA:
-	adds r2, r0
-	mov r12, r2
-	mov r2, r9
-	ldrh r0, [r2, 0x6]
-	cmp r12, r0
-	ble _08004CC8
-	mov r12, r0
-_08004CC8:
-	movs r0, 0x7
-	ands r0, r1
-	adds r0, r1, r0
-	asrs r0, 3
-	str r0, [sp]
-	adds r1, r5, 0
-	cmp r1, r12
-	bge _08004D30
-_08004CD8:
-	mov r3, r10
-	adds r0, r1, 0x1
-	mov r8, r0
-	cmp r3, r4
-	bge _08004D2A
-	asrs r0, r1, 3
-	ldr r2, [sp]
-	muls r0, r2
-	lsls r6, r0, 5
-	lsls r0, r1, 29
-	lsrs r5, r0, 27
-_08004CEE:
-	asrs r2, r3, 1
-	movs r0, 0x3
-	ands r2, r0
-	mov r0, r9
-	ldr r1, [r0]
-	adds r1, r2
-	asrs r0, r3, 3
-	lsls r0, 5
-	adds r1, r0
-	adds r1, r6
-	adds r2, r1, r5
-	movs r0, 0x1
-	ands r0, r3
-	cmp r0, 0
-	beq _08004D1A
-	ldrb r0, [r2]
-	movs r1, 0xF
-	ands r1, r0
-	lsls r0, r7, 4
-	orrs r1, r0
-	strb r1, [r2]
-	b _08004D24
-_08004D1A:
-	ldrb r1, [r2]
-	movs r0, 0xF0
-	ands r0, r1
-	orrs r0, r7
-	strb r0, [r2]
-_08004D24:
-	adds r3, 0x1
-	cmp r3, r4
-	blt _08004CEE
-_08004D2A:
-	mov r1, r8
-	cmp r1, r12
-	blt _08004CD8
-_08004D30:
-	add sp, 0x4
-	pop {r3-r5}
-	mov r8, r3
-	mov r9, r4
-	mov r10, r5
-	pop {r4-r7}
-	pop {r0}
-	bx r0
-	thumb_func_end FillBitmapRect4Bit
-
-	thumb_func_start BlitBitmapRect4BitTo8Bit
-BlitBitmapRect4BitTo8Bit: @ 8004D40
-	push {r4-r7,lr}
-	mov r7, r10
-	mov r6, r9
-	mov r5, r8
-	push {r5-r7}
-	sub sp, 0x3C
-	str r0, [sp]
-	str r1, [sp, 0x4]
-	ldr r0, [sp, 0x5C]
-	ldr r1, [sp, 0x60]
-	ldr r5, [sp, 0x64]
-	ldr r4, [sp, 0x68]
-	ldr r6, [sp, 0x6C]
-	ldr r7, [sp, 0x70]
-	mov r8, r7
-	lsls r2, 16
-	lsrs r2, 16
-	str r2, [sp, 0x8]
-	lsls r3, 16
-	lsrs r3, 16
-	lsls r0, 16
-	lsrs r0, 16
-	str r0, [sp, 0xC]
-	lsls r1, 16
-	lsrs r7, r1, 16
-	lsls r5, 16
-	lsrs r5, 16
-	lsls r4, 16
-	lsrs r4, 16
-	lsls r6, 24
-	lsrs r6, 24
-	str r6, [sp, 0x10]
-	mov r0, r8
-	lsls r0, 28
-	lsrs r0, 24
-	str r0, [sp, 0x14]
-	lsls r0, r6, 28
-	lsrs r0, 24
-	str r0, [sp, 0x28]
-	ldr r1, [sp, 0x4]
-	ldrh r0, [r1, 0x4]
-	ldr r2, [sp, 0xC]
-	subs r0, r2
-	ldr r1, [sp, 0x8]
-	adds r1, r5, r1
-	str r1, [sp, 0x18]
-	cmp r0, r5
-	bge _08004DA6
-	ldr r2, [sp, 0x8]
-	adds r0, r2
-	str r0, [sp, 0x18]
-_08004DA6:
-	ldr r5, [sp, 0x4]
-	ldrh r1, [r5, 0x6]
-	subs r0, r1, r7
-	cmp r0, r4
-	bge _08004DB8
-	adds r0, r3, r1
-	subs r0, r7
-	str r0, [sp, 0x1C]
-	b _08004DBC
-_08004DB8:
-	adds r4, r3, r4
-	str r4, [sp, 0x1C]
-_08004DBC:
-	ldr r0, [sp]
-	ldrh r1, [r0, 0x4]
-	movs r2, 0x7
-	adds r0, r1, 0
-	ands r0, r2
-	adds r1, r0
-	asrs r1, 3
-	str r1, [sp, 0x20]
-	ldr r5, [sp, 0x4]
-	ldrh r1, [r5, 0x4]
-	adds r0, r1, 0
-	ands r0, r2
-	adds r1, r0
-	asrs r1, 3
-	str r1, [sp, 0x24]
-	ldr r0, [sp, 0x10]
-	cmp r0, 0xFF
-	bne _08004E9C
-	adds r2, r3, 0
-	adds r5, r7, 0
-	ldr r1, [sp, 0x1C]
-	cmp r2, r1
-	blt _08004DEC
-	b _08004F7E
-_08004DEC:
-	ldr r3, [sp, 0x8]
-	lsrs r3, 1
-	str r3, [sp, 0x2C]
-	movs r0, 0x3
-	ands r3, r0
-	str r3, [sp, 0x2C]
-	ldr r7, [sp, 0x8]
-	lsrs r0, r7, 3
-	lsls r0, 5
-	str r0, [sp, 0x30]
-_08004E00:
-	ldr r0, [sp]
-	ldr r1, [r0]
-	ldr r3, [sp, 0x2C]
-	adds r1, r3
-	ldr r7, [sp, 0x30]
-	adds r1, r7
-	asrs r0, r2, 3
-	ldr r3, [sp, 0x20]
-	muls r0, r3
-	lsls r0, 5
-	mov r8, r0
-	add r1, r8
-	lsls r0, r2, 29
-	lsrs r7, r0, 27
-	adds r6, r1, r7
-	ldr r3, [sp, 0x8]
-	ldr r4, [sp, 0xC]
-	adds r2, 0x1
-	mov r10, r2
-	adds r0, r5, 0x1
-	str r0, [sp, 0x38]
-	ldr r1, [sp, 0x18]
-	cmp r3, r1
-	bge _08004E90
-	asrs r0, r5, 3
-	ldr r2, [sp, 0x24]
-	muls r0, r2
-	lsls r0, 6
-	mov r12, r0
-	lsls r0, r5, 29
-	lsrs r0, 26
-	mov r9, r0
-_08004E40:
-	movs r0, 0x7
-	ands r0, r4
-	ldr r5, [sp, 0x4]
-	ldr r1, [r5]
-	adds r1, r0
-	asrs r0, r4, 3
-	lsls r0, 6
-	adds r1, r0
-	add r1, r12
-	mov r0, r9
-	adds r5, r1, r0
-	movs r0, 0x1
-	ands r0, r3
-	cmp r0, 0
-	beq _08004E64
-	ldrb r0, [r6]
-	lsrs r0, 4
-	b _08004E80
-_08004E64:
-	asrs r2, r3, 1
-	movs r0, 0x3
-	ands r2, r0
-	ldr r0, [sp]
-	ldr r1, [r0]
-	adds r1, r2
-	asrs r0, r3, 3
-	lsls r0, 5
-	adds r1, r0
-	add r1, r8
-	adds r6, r1, r7
-	ldrb r1, [r6]
-	movs r0, 0xF
-	ands r0, r1
-_08004E80:
-	ldr r1, [sp, 0x14]
-	adds r0, r1, r0
-	strb r0, [r5]
-	adds r3, 0x1
-	adds r4, 0x1
-	ldr r2, [sp, 0x18]
-	cmp r3, r2
-	blt _08004E40
-_08004E90:
-	mov r2, r10
-	ldr r5, [sp, 0x38]
-	ldr r3, [sp, 0x1C]
-	cmp r2, r3
-	blt _08004E00
-	b _08004F7E
-_08004E9C:
-	adds r2, r3, 0
-	adds r5, r7, 0
-	ldr r7, [sp, 0x1C]
-	cmp r2, r7
-	bge _08004F7E
-	ldr r0, [sp, 0x8]
-	lsrs r0, 1
-	str r0, [sp, 0x34]
-	movs r0, 0x3
-	ldr r1, [sp, 0x34]
-	ands r1, r0
-	str r1, [sp, 0x34]
-_08004EB4:
-	ldr r3, [sp]
-	ldr r1, [r3]
-	ldr r7, [sp, 0x34]
-	adds r1, r7
-	ldr r3, [sp, 0x8]
-	lsrs r0, r3, 3
-	lsls r0, 5
-	adds r1, r0
-	asrs r0, r2, 3
-	ldr r7, [sp, 0x20]
-	muls r0, r7
-	lsls r0, 5
-	mov r8, r0
-	add r1, r8
-	lsls r0, r2, 29
-	lsrs r7, r0, 27
-	adds r6, r1, r7
-	ldr r4, [sp, 0xC]
-	adds r2, 0x1
-	mov r10, r2
-	adds r0, r5, 0x1
-	str r0, [sp, 0x38]
-	ldr r1, [sp, 0x18]
-	cmp r3, r1
-	bge _08004F74
-	asrs r0, r5, 3
-	ldr r2, [sp, 0x24]
-	muls r0, r2
-	lsls r0, 6
-	mov r9, r0
-	lsls r0, r5, 29
-	lsrs r0, 26
-	mov r12, r0
-_08004EF6:
-	movs r0, 0x1
-	ands r0, r3
-	cmp r0, 0
-	beq _08004F2A
-	ldrb r2, [r6]
-	movs r0, 0xF0
-	ands r0, r2
-	ldr r5, [sp, 0x28]
-	cmp r0, r5
-	beq _08004F6A
-	adds r0, r4, 0
-	movs r1, 0x7
-	ands r0, r1
-	ldr r5, [sp, 0x4]
-	ldr r1, [r5]
-	adds r1, r0
-	asrs r0, r4, 3
-	lsls r0, 6
-	adds r1, r0
-	add r1, r9
-	mov r0, r12
-	adds r5, r1, r0
-	lsrs r0, r2, 4
-	ldr r1, [sp, 0x14]
-	adds r0, r1, r0
-	b _08004F68
-_08004F2A:
-	asrs r2, r3, 1
-	movs r0, 0x3
-	ands r2, r0
-	ldr r5, [sp]
-	ldr r1, [r5]
-	adds r1, r2
-	asrs r0, r3, 3
-	lsls r0, 5
-	adds r1, r0
-	add r1, r8
-	adds r6, r1, r7
-	ldrb r0, [r6]
-	movs r2, 0xF
-	ands r2, r0
-	ldr r0, [sp, 0x10]
-	cmp r2, r0
-	beq _08004F6A
-	adds r0, r4, 0
-	movs r1, 0x7
-	ands r0, r1
-	ldr r5, [sp, 0x4]
-	ldr r1, [r5]
-	adds r1, r0
-	asrs r0, r4, 3
-	lsls r0, 6
-	adds r1, r0
-	add r1, r9
-	mov r0, r12
-	adds r5, r1, r0
-	ldr r1, [sp, 0x14]
-	adds r0, r1, r2
-_08004F68:
-	strb r0, [r5]
-_08004F6A:
-	adds r3, 0x1
-	adds r4, 0x1
-	ldr r2, [sp, 0x18]
-	cmp r3, r2
-	blt _08004EF6
-_08004F74:
-	mov r2, r10
-	ldr r5, [sp, 0x38]
-	ldr r3, [sp, 0x1C]
-	cmp r2, r3
-	blt _08004EB4
-_08004F7E:
-	add sp, 0x3C
-	pop {r3-r5}
-	mov r8, r3
-	mov r9, r4
-	mov r10, r5
-	pop {r4-r7}
-	pop {r0}
-	bx r0
-	thumb_func_end BlitBitmapRect4BitTo8Bit
-
-	thumb_func_start FillBitmapRect8Bit
-FillBitmapRect8Bit: @ 8004F90
-	push {r4-r7,lr}
-	mov r7, r10
-	mov r6, r9
-	mov r5, r8
-	push {r5-r7}
-	sub sp, 0x8
-	mov r8, r0
-	ldr r0, [sp, 0x28]
-	ldr r4, [sp, 0x2C]
-	lsls r1, 16
-	lsrs r1, 16
-	mov r10, r1
-	lsls r2, 16
-	lsrs r2, 16
-	adds r6, r2, 0
-	lsls r3, 16
-	lsrs r3, 16
-	lsls r0, 16
-	lsrs r0, 16
-	lsls r4, 24
-	lsrs r4, 24
-	mov r9, r4
-	adds r5, r1, r3
-	mov r3, r8
-	ldrh r1, [r3, 0x4]
-	cmp r5, r1
-	ble _08004FC8
-	adds r5, r1, 0
-_08004FC8:
-	adds r7, r2, r0
-	mov r2, r8
-	ldrh r0, [r2, 0x6]
-	cmp r7, r0
-	ble _08004FD4
-	adds r7, r0, 0
-_08004FD4:
-	movs r0, 0x7
-	ands r0, r1
-	adds r0, r1, r0
-	asrs r0, 3
-	str r0, [sp]
-	adds r1, r6, 0
-	cmp r1, r7
-	bge _08005024
-_08004FE4:
-	mov r3, r10
-	adds r6, r1, 0x1
-	str r6, [sp, 0x4]
-	cmp r3, r5
-	bge _0800501E
-	movs r0, 0x7
-	mov r12, r0
-	asrs r0, r1, 3
-	ldr r2, [sp]
-	muls r0, r2
-	lsls r4, r0, 6
-	lsls r0, r1, 29
-	lsrs r2, r0, 26
-_08004FFE:
-	adds r0, r3, 0
-	mov r6, r12
-	ands r0, r6
-	mov r6, r8
-	ldr r1, [r6]
-	adds r1, r0
-	asrs r0, r3, 3
-	lsls r0, 6
-	adds r1, r0
-	adds r1, r4
-	adds r1, r2
-	mov r0, r9
-	strb r0, [r1]
-	adds r3, 0x1
-	cmp r3, r5
-	blt _08004FFE
-_0800501E:
-	ldr r1, [sp, 0x4]
-	cmp r1, r7
-	blt _08004FE4
-_08005024:
-	add sp, 0x8
-	pop {r3-r5}
-	mov r8, r3
-	mov r9, r4
-	mov r10, r5
-	pop {r4-r7}
-	pop {r0}
-	bx r0
-	thumb_func_end FillBitmapRect8Bit
-
-	.align 2, 0 @ Don't pad with nop.
diff --git a/include/blit.h b/include/blit.h
new file mode 100644
index 000000000..78f67766e
--- /dev/null
+++ b/include/blit.h
@@ -0,0 +1,17 @@
+#ifndef GUARD_BLIT_H
+#define GUARD_BLIT_H
+
+struct Bitmap
+{
+    u8 *pixels;
+    u32 width:16;
+    u32 height:16;
+};
+
+void BlitBitmapRect4BitWithoutColorKey(const struct Bitmap *src, struct Bitmap *dst, u16 srcX, u16 srcY, u16 dstX, u16 dstY, u16 width, u16 height);
+void BlitBitmapRect4Bit(const struct Bitmap *src, struct Bitmap *dst, u16 srcX, u16 srcY, u16 dstX, u16 dstY, u16 width, u16 height, u8 colorKey);
+void FillBitmapRect4Bit(struct Bitmap *surface, u16 x, u16 y, u16 width, u16 height, u8 fillValue);
+void BlitBitmapRect4BitTo8Bit(const struct Bitmap *src, struct Bitmap *dst, u16 srcX, u16 srcY, u16 dstX, u16 dstY, u16 width, u16 height, u8 colorKey, u8 paletteOffset);
+void FillBitmapRect8Bit(struct Bitmap *surface, u16 x, u16 y, u16 width, u16 height, u8 fillValue);
+
+#endif // GUARD_BLIT_H
diff --git a/include/global.h b/include/global.h
index 8e85b3123..de2221f8c 100644
--- a/include/global.h
+++ b/include/global.h
@@ -760,13 +760,6 @@ struct SaveBlock1
 
 extern struct SaveBlock1* gSaveBlock1Ptr;
 
-struct Bitmap           // TODO: Find a better spot for this
-{
-    u8* pixels;
-    u32 width:16;
-    u32 height:16;
-};
-
 extern u8 gReservedSpritePaletteCount;
 
 #endif // GUARD_GLOBAL_H
diff --git a/ld_script.txt b/ld_script.txt
index 88c2c59bc..53385c39f 100644
--- a/ld_script.txt
+++ b/ld_script.txt
@@ -55,7 +55,7 @@ SECTIONS {
         src/text_printer.o(.text);
         asm/text_printer.o(.text);
         src/window.o(.text);
-        asm/blit.o(.text);
+        src/blit.o(.text);
         src/window_8bpp.o(.text);
         src/text.o(.text);
         src/sprite.o(.text);
diff --git a/src/blit.c b/src/blit.c
new file mode 100644
index 000000000..969f89900
--- /dev/null
+++ b/src/blit.c
@@ -0,0 +1,212 @@
+#include "global.h"
+#include "blit.h"
+
+void BlitBitmapRect4BitWithoutColorKey(const struct Bitmap *src, struct Bitmap *dst, u16 srcX, u16 srcY, u16 dstX, u16 dstY, u16 width, u16 height)
+{
+    BlitBitmapRect4Bit(src, dst, srcX, srcY, dstX, dstY, width, height, 0xFF);
+}
+
+void BlitBitmapRect4Bit(const struct Bitmap *src, struct Bitmap *dst, u16 srcX, u16 srcY, u16 dstX, u16 dstY, u16 width, u16 height, u8 colorKey)
+{
+    s32 xEnd;
+    s32 yEnd;
+    s32 multiplierSrcY;
+    s32 multiplierDstY;
+    s32 loopSrcY, loopDstY;
+    s32 loopSrcX, loopDstX;
+    const u8 *pixelsSrc;
+    u8 *pixelsDst;
+    s32 toOrr;
+    s32 toAnd;
+    s32 toShift;
+
+    if (dst->width - dstX < width)
+        xEnd = (dst->width - dstX) + srcX;
+    else
+        xEnd = srcX + width;
+
+    if (dst->height - dstY < height)
+        yEnd = (dst->height - dstY) + srcY;
+    else
+        yEnd = height + srcY;
+
+    multiplierSrcY = (src->width + (src->width & 7)) >> 3;
+    multiplierDstY = (dst->width + (dst->width & 7)) >> 3;
+
+    if (colorKey == 0xFF)
+    {
+        for (loopSrcY = srcY, loopDstY = dstY; loopSrcY < yEnd; loopSrcY++, loopDstY++)
+        {
+            for (loopSrcX = srcX, loopDstX = dstX; loopSrcX < xEnd; loopSrcX++, loopDstX++)
+            {
+                pixelsSrc = src->pixels + ((loopSrcX >> 1) & 3) + ((loopSrcX >> 3) << 5) + (((loopSrcY >> 3) * multiplierSrcY) << 5) + ((u32)(loopSrcY << 0x1d) >> 0x1B);
+                pixelsDst = dst->pixels + ((loopDstX >> 1) & 3) + ((loopDstX >> 3) << 5) + (((loopDstY >> 3) * multiplierDstY) << 5) + ((u32)(loopDstY << 0x1d) >> 0x1B);
+                toOrr = ((*pixelsSrc >> ((loopSrcX & 1) << 2)) & 0xF);
+                toShift = ((loopDstX & 1) << 2);
+                toOrr <<= toShift;
+                toAnd = 0xF0 >> (toShift);
+                *pixelsDst = toOrr | (*pixelsDst & toAnd);
+            }
+        }
+    }
+    else
+    {
+        for (loopSrcY = srcY, loopDstY = dstY; loopSrcY < yEnd; loopSrcY++, loopDstY++)
+        {
+            for (loopSrcX = srcX, loopDstX = dstX; loopSrcX < xEnd; loopSrcX++, loopDstX++)
+            {
+                pixelsSrc = src->pixels + ((loopSrcX >> 1) & 3) + ((loopSrcX >> 3) << 5) + (((loopSrcY >> 3) * multiplierSrcY) << 5) + ((u32)(loopSrcY << 0x1d) >> 0x1B);
+                pixelsDst = dst->pixels + ((loopDstX >> 1) & 3) + ((loopDstX >> 3) << 5) + (((loopDstY >> 3) * multiplierDstY) << 5) + ((u32)(loopDstY << 0x1d) >> 0x1B);
+                toOrr = ((*pixelsSrc >> ((loopSrcX & 1) << 2)) & 0xF);
+                if (toOrr != colorKey)
+                {
+                    toShift = ((loopDstX & 1) << 2);
+                    toOrr <<= toShift;
+                    toAnd = 0xF0 >> (toShift);
+                    *pixelsDst = toOrr | (*pixelsDst & toAnd);
+                }
+            }
+        }
+    }
+}
+
+void FillBitmapRect4Bit(struct Bitmap *surface, u16 x, u16 y, u16 width, u16 height, u8 fillValue)
+{
+    s32 xEnd;
+    s32 yEnd;
+    s32 multiplierY;
+    s32 loopX, loopY;
+
+    xEnd = x + width;
+    if (xEnd > surface->width)
+        xEnd = surface->width;
+
+    yEnd = y + height;
+    if (yEnd > surface->height)
+        yEnd = surface->height;
+
+    multiplierY = (surface->width + (surface->width & 7)) >> 3;
+
+    for (loopY = y; loopY < yEnd; loopY++)
+    {
+        for (loopX = x; loopX < xEnd; loopX++)
+        {
+            u8 *pixels = surface->pixels + ((loopX >> 1) & 3) + ((loopX >> 3) << 5) + (((loopY >> 3) * multiplierY) << 5) + ((u32)(loopY << 0x1d) >> 0x1B);
+            if ((loopX & 1) != 0)
+            {
+                *pixels &= 0xF;
+                *pixels |= fillValue << 4;
+            }
+            else
+            {
+                *pixels &= 0xF0;
+                *pixels |= fillValue;
+            }
+        }
+    }
+}
+
+void BlitBitmapRect4BitTo8Bit(const struct Bitmap *src, struct Bitmap *dst, u16 srcX, u16 srcY, u16 dstX, u16 dstY, u16 width, u16 height, u8 colorKey, u8 paletteOffset)
+{
+    s32 palOffsetBits;
+    s32 xEnd;
+    s32 yEnd;
+    s32 multiplierSrcY;
+    s32 multiplierDstY;
+    s32 loopSrcY, loopDstY;
+    s32 loopSrcX, loopDstX;
+    const u8 *pixelsSrc;
+    u8 *pixelsDst;
+    s32 colorKeyBits;
+
+    palOffsetBits = (u32)(paletteOffset << 0x1C) >> 0x18;
+    colorKeyBits = (u32)(colorKey << 0x1C) >> 0x18;
+
+    if (dst->width - dstX < width)
+        xEnd = (dst->width - dstX) + srcX;
+    else
+        xEnd = width + srcX;
+
+    if (dst->height - dstY < height)
+        yEnd = (srcY + dst->height) - dstY;
+    else
+        yEnd = srcY + height;
+
+    multiplierSrcY = (src->width + (src->width & 7)) >> 3;
+    multiplierDstY = (dst->width + (dst->width & 7)) >> 3;
+
+    if (colorKey == 0xFF)
+    {
+        for (loopSrcY = srcY, loopDstY = dstY; loopSrcY < yEnd; loopSrcY++, loopDstY++)
+        {
+            pixelsSrc = src->pixels + ((srcX >> 1) & 3) + ((srcX >> 3) << 5) + (((loopSrcY >> 3) * multiplierSrcY) << 5) + ((u32)(loopSrcY << 0x1d) >> 0x1b);
+            for (loopSrcX = srcX, loopDstX = dstX; loopSrcX < xEnd; loopSrcX++, loopDstX++)
+            {
+                pixelsDst = dst->pixels + (loopDstX & 7) + ((loopDstX >> 3) << 6) + (((loopDstY >> 3) * multiplierDstY) << 6) + ((u32)(loopDstY << 0x1d) >> 0x1a);
+                if (loopSrcX & 1)
+                {
+                    *pixelsDst = palOffsetBits + (*pixelsSrc >> 4);
+                }
+                else
+                {
+                    pixelsSrc = src->pixels + ((loopSrcX >> 1) & 3) + ((loopSrcX >> 3) << 5) + (((loopSrcY >> 3) * multiplierSrcY) << 5) + ((u32)(loopSrcY << 0x1d) >> 0x1b);
+                    *pixelsDst = palOffsetBits + (*pixelsSrc & 0xF);
+                }
+            }
+        }
+    }
+    else
+    {
+        for (loopSrcY = srcY, loopDstY = dstY; loopSrcY < yEnd; loopSrcY++, loopDstY++)
+        {
+            pixelsSrc = src->pixels + ((srcX >> 1) & 3) + ((srcX >> 3) << 5) + (((loopSrcY >> 3) * multiplierSrcY) << 5) + ((u32)(loopSrcY << 0x1d) >> 0x1b);
+            for (loopSrcX = srcX, loopDstX = dstX; loopSrcX < xEnd; loopSrcX++, loopDstX++)
+            {
+                if (loopSrcX & 1)
+                {
+                    if ((*pixelsSrc & 0xF0) != colorKeyBits)
+                    {
+                        pixelsDst = dst->pixels + (loopDstX & 7) + ((loopDstX >> 3) << 6) + (((loopDstY >> 3) * multiplierDstY) << 6) + ((u32)(loopDstY << 0x1d) >> 0x1a);
+                        *pixelsDst = palOffsetBits + (*pixelsSrc >> 4);
+                    }
+                }
+                else
+                {
+                    pixelsSrc = src->pixels + ((loopSrcX >> 1) & 3) + ((loopSrcX >> 3) << 5) + (((loopSrcY >> 3) * multiplierSrcY) << 5) + ((u32)(loopSrcY << 0x1d) >> 0x1b);
+                    if ((*pixelsSrc & 0xF) != colorKey)
+                    {
+                        pixelsDst = dst->pixels + (loopDstX & 7) + ((loopDstX >> 3) << 6) + (((loopDstY >> 3) * multiplierDstY) << 6) + ((u32)(loopDstY << 0x1d) >> 0x1a);
+                        *pixelsDst = palOffsetBits + (*pixelsSrc & 0xF);
+                    }
+                }
+            }
+        }
+    }
+}
+
+void FillBitmapRect8Bit(struct Bitmap *surface, u16 x, u16 y, u16 width, u16 height, u8 fillValue)
+{
+    s32 xEnd;
+    s32 yEnd;
+    s32 multiplierY;
+    s32 loopX, loopY;
+
+    xEnd = x + width;
+    if (xEnd > surface->width)
+        xEnd = surface->width;
+
+    yEnd = y + height;
+    if (yEnd > surface->height)
+        yEnd = surface->height;
+
+    multiplierY = (surface->width + (surface->width & 7)) >> 3;
+
+    for (loopY = y; loopY < yEnd; loopY++)
+    {
+        for (loopX = x; loopX < xEnd; loopX++)
+        {
+            u8 *pixels = surface->pixels + (loopX & 7) + ((loopX >> 3) << 6) + (((loopY >> 3) * multiplierY) << 6) + ((u32)(loopY << 0x1d) >> 0x1a);
+            *pixels = fillValue;
+        }
+    }
+}
diff --git a/src/text.c b/src/text.c
index d74d77e2f..34ceb92d4 100644
--- a/src/text.c
+++ b/src/text.c
@@ -5,13 +5,13 @@
 #include "window.h"
 #include "text.h"
 #include "sprite.h"
+#include "blit.h"
 
 extern u8 gGlyphInfo[0x90];
 extern u8 gUnknown_203ADFA;
 extern u16 gTMCaseMainWindowPalette[];
 extern const struct OamData gOamData_83AC9D0;
 
-extern void FillBitmapRect4Bit(struct Bitmap *surface, u16 x, u16 y, u16 width, u16 height, u8 fillValue);
 extern void FillWindowPixelRect(u8 windowId, u8 fillValue, u16 x, u16 y, u16 width, u16 height);
 extern void BlitBitmapRectToWindow(u8 windowId, const u8 *pixels, u16 srcX, u16 srcY, u16 srcWidth, int srcHeight, u16 destX, u16 destY, u16 rectWidth, u16 rectHeight);
 extern u8 GetKeypadIconWidth(u8 keypadIconId);
diff --git a/src/window.c b/src/window.c
index 71cfa6898..c4ea49956 100644
--- a/src/window.c
+++ b/src/window.c
@@ -2,17 +2,13 @@
 #include "window.h"
 #include "malloc.h"
 #include "bg.h"
+#include "blit.h"
 
 u8 gWindowClearTile;
 void *gWindowBgTilemapBuffers[4];
 
 EWRAM_DATA struct Window gWindows[WINDOWS_MAX] = {0};
 
-extern void BlitBitmapRect4Bit(struct Bitmap *src, struct Bitmap *dest, u16 srcX, u16 srcY, u16 destX, u16 destY, u16 width, u16 height, u8 colorKey);
-extern void BlitBitmapRect4BitTo8Bit(struct Bitmap *src, struct Bitmap *dest, u16 srcX, u16 srcY, u16 destX, u16 destY, u16 width, u16 height, u8 colorKey, u8 paletteNum);
-extern void FillBitmapRect4Bit(struct Bitmap *surface, u16 x, u16 y, u16 width, u16 height, u8 fillValue);
-extern void FillBitmapRect8Bit(struct Bitmap *surface, u16 x, u16 y, u16 width, u16 height, u8 fillValue);
-
 static u8 GetNumActiveWindowsOnBg(u8 bgId);
 
 static const struct WindowTemplate sDummyWindowTemplate = {0xFF, 0, 0, 0, 0, 0, 0};
diff --git a/src/window_8bpp.c b/src/window_8bpp.c
index 5eac3c558..8e5d19565 100644
--- a/src/window_8bpp.c
+++ b/src/window_8bpp.c
@@ -2,13 +2,11 @@
 #include "window.h"
 #include "malloc.h"
 #include "bg.h"
+#include "blit.h"
 
 EWRAM_DATA static struct Window* sWindowPtr = NULL;
 EWRAM_DATA static u16 sWindowSize = 0;
 
-extern void BlitBitmapRect4BitTo8Bit(struct Bitmap *src, struct Bitmap *dest, u16 srcX, u16 srcY, u16 destX, u16 destY, u16 width, u16 height, u8 colorKey, u8 paletteNum);
-extern void FillBitmapRect8Bit(struct Bitmap *surface, u16 x, u16 y, u16 width, u16 height, u8 fillValue);
-
 static u8 GetNumActiveWindowsOnBg8Bit(u8 bgId);
 
 static void nullsub_9(void)