[Decompilation] [th05] .CDG: Non-color blitting

And this is how you make code less undecompilable by improving your
pointless micro-optimizations to use more registers instead of
self-modifying code. Worth it if only to get rid of the branches in
TH04's undecompilable ASM implementation.

Part of P0133, funded by [Anonymous].

Makefile.mak

@@ -150,7 +150,7 @@ bin\th05\res_kso.com: th05\res_kso.cpp
 $**
 | masters.lib
 
-bin\th05\op.exe: th05\op010.cpp bin\th05\op.obj th05\op011.cpp th05\m_char.cpp bin\th05\frmdelay.obj bin\th04\cdg_load.obj bin\th05\egcrect.obj bin\hfliplut.obj
+bin\th05\op.exe: th05\op010.cpp bin\th05\op.obj th05\op011.cpp th05\m_char.cpp th05\cdg_p_nc.cpp bin\th05\frmdelay.obj bin\th04\cdg_load.obj bin\th05\egcrect.obj bin\hfliplut.obj
 	$(CC) $(CFLAGS) -ml -DGAME=5 -DBINARY='O' -3 -Z -nbin\th05\ -eOP.EXE @&&|
 $**
 |

th04/formats/cdg_put_nocolors.asm

@@ -1,3 +1,5 @@
+; Identical to the barely decompilable TH05 version, which doesn't rely on
+; self-modifying code for the width.
 public CDG_PUT_NOCOLORS_8
 cdg_put_nocolors_8 proc far
 
@@ -16,41 +18,25 @@ cdg_put_nocolors_8 proc far
 	sar	di, 3
 	add	di, [si+cdg_t.offset_at_bottom_left]
 	mov	ax, [si+cdg_t.vram_dword_w]
-if GAME eq 4
 	mov	word ptr cs:@@width+1, ax
 	jmp	short $+2
-else
-	mov	bx, ax
-endif
 	shl	ax, 2
 	add	ax, (640 / 8)
 	mov	dx, ax
 	mov	ax, [bp+@@top]
-if GAME eq 4
 	mov	bx, ax
 	shl	ax, 2
 	add	ax, bx
-else
-	mov	cx, ax
-	shl	ax, 2
-	add	ax, cx
-endif
 	add	ax, 0A800h
 	mov	es, ax
 	push	ds
 	mov	ds, [si+cdg_t.seg_alpha]
 	xor	si, si
-if GAME eq 4
 	cld
-endif
 	align 2
 
 @@width:
-if GAME eq 4
 	mov	cx, 1234h
-else
-	mov	cx, bx
-endif
 	rep movsd
 	sub	di, dx
 	jns	short @@width

th05/cdg_p_nc.cpp

@@ -0,0 +1 @@
+#include "th05/formats/cdg_p_nc.cpp"

th05/formats/cdg_p_nc.cpp

@@ -0,0 +1,61 @@
+#pragma codeseg SHARED_
+
+extern "C" {
+#include <stddef.h>
+#include <dos.h>
+#include "platform.h"
+#include "pc98.h"
+#include "planar.h"
+#include "decomp.h"
+#include "th04/formats/cdg.h"
+}
+
+void pascal cdg_put_nocolors_8(screen_x_t left, vram_y_t top, int slot)
+{
+	#define _DI	static_cast<vram_offset_t>(_DI)
+	#define cdg	reinterpret_cast<cdg_t near *>(_SI)
+	#define stride	static_cast<vram_byte_amount_t>(_DX)
+	#define tmp	static_cast<upixel_t>(_AX)
+	#define top_tmp	static_cast<pixel_t>(_CX)
+	register vram_byte_amount_t cdg_dword_w;
+
+	// cdg = cdg_slots[slot];
+	_SI = slot;
+	_SI <<= 4;	// MODDERS: Should be sizeof(cdg_t);
+	_SI += FP_OFF(cdg_slots);
+
+	// _DI = (vram_offset_shift(left, 0) + cdg->offset_at_bottom_left);
+	// [top] is calculated into the segment later, not the offset.
+	_DI = left;
+	_DI >>= 3;
+	_DI += cdg->offset_at_bottom_left;
+
+	tmp = cdg->vram_dword_w;
+	cdg_dword_w = tmp;
+	// One additional full line, since we move from the end of one CDG line to
+	// the beginning of the previous one.
+	stride = ((tmp * 4) + ROW_SIZE);
+
+	// _ES = (SEG_PLANE_B + (top * (ROW_SIZE / 16)));
+	tmp = top;
+	top_tmp = tmp;
+	_ES = (((tmp * 4) + top_tmp) + SEG_PLANE_B);
+
+	__asm { push ds };
+	_DS = reinterpret_cast<seg_t>(cdg->seg_alpha());
+	_SI = 0;
+
+	do {
+		_CX = cdg_dword_w;
+		REP MOVSD;
+		_DI -= stride;
+	} while(!FLAGS_SIGN);
+
+	__asm { pop ds };
+
+	#undef top_tmp
+	#undef tmp
+	#undef stride
+	#undef cdg
+	#undef _DI
+}

th05_op.asm

@@ -2554,7 +2554,7 @@ include th05/hardware/input_held.asm
 include th05/hardware/input_wait.asm
 include th05/snd/measure.asm
 include th05/snd/delaymea.asm
-include th04/formats/cdg_put_nocolors.asm
+	extern CDG_PUT_NOCOLORS_8:proc
 	extern FRAME_DELAY:proc
 	extern CDG_LOAD_SINGLE_NOALPHA:proc
 	extern CDG_LOAD_SINGLE:proc