From 15033b27a4926212305e00513bc98c5d18d55074 Mon Sep 17 00:00:00 2001
From: Sjoerd van der Berg <harekiet@users.sourceforge.net>
Date: Fri, 25 Oct 2002 21:25:35 +0000
Subject: [PATCH] New prefix variables and a new string repeat function.

Imported-from: https://svn.code.sf.net/p/dosbox/code-0/dosbox/trunk@427
---
 src/cpu/core_16/support.h | 317 +++++++++++++++++++++++++++++++++-----
 1 file changed, 276 insertions(+), 41 deletions(-)

diff --git a/src/cpu/core_16/support.h b/src/cpu/core_16/support.h
index 095f585a..ace3a10f 100644
--- a/src/cpu/core_16/support.h
+++ b/src/cpu/core_16/support.h
@@ -88,8 +88,8 @@ static INLINE Bit32u Pop_32() {
 
 #define stringSI											\
 	EAPoint from;											\
-	if (prefixes & PREFIX_SEG) {										\
-		from=(segprefix_base+reg_si);						\
+	if (prefix.mark & PREFIX_SEG) {										\
+		from=(prefix.segbase+reg_si);						\
 		PrefixReset;										\
 	} else {												\
 		from=SegBase(ds)+reg_si;							\
@@ -102,48 +102,265 @@ static INLINE Bit32u Pop_32() {
 #include "instructions.h"
 
 
-static INLINE void Rep_66(Bit16s direct,EAPoint from,EAPoint to) {
-	bool again;
-	do {
-		again=false;
-		Bit8u repcode=Fetchb();
-		switch (repcode) {
-		case 0x26:			/* ES Prefix */
-			again=true;
-			from=SegBase(es);
-			break;
-		case 0x2e:			/* CS Prefix */
-			again=true;
-			from=SegBase(cs);
-			break;
-		case 0x36:			/* SS Prefix */
-			again=true;
-			from=SegBase(ss);
-			break;
-		case 0x3e:			/* DS Prefix */
-			again=true;
-			from=SegBase(ds);
-			break;
-		case 0xa5:			/* REP MOVSD */
-			for (;reg_cx>0;reg_cx--) {
-				SaveMd(to+reg_di,LoadMd(from+reg_si));
-				reg_di+=direct*4;
-				reg_si+=direct*4;
-			}	
-			break;
-		case 0xab:			/* REP STOSW */
-			for (;reg_cx>0;reg_cx--) {
-				SaveMd(to+reg_di,reg_eax);
-				reg_di+=direct*4;
-			}	
-			break;
-		default:
-			E_Exit("CPU:Opcode 66:Illegal REP prefix %2X",repcode);
+static Bits Repeat_Normal(bool testz,bool prefix_66,Bits count_remain) {
+	
+	Bits count=count_remain;
+	PhysPt base_si,base_di;
+
+	Bit16s direct;
+	if (flags.df) direct=-1;
+	else direct=1;
+	base_di=SegBase(es);
+rep_again:	
+	if (prefix.mark & PREFIX_SEG) {
+		base_si=(prefix.segbase);
+	} else {
+		base_si=SegBase(ds);
+	}
+	switch (Fetchb()) {
+	case 0x26:			/* ES Prefix */
+		prefix.segbase=SegBase(es);
+		prefix.mark|=PREFIX_SEG;
+		prefix.count++;
+		goto rep_again;
+	case 0x2e:			/* CS Prefix */
+		prefix.segbase=SegBase(cs);
+		prefix.mark|=PREFIX_SEG;
+		prefix.count++;
+		goto rep_again;
+	case 0x36:			/* SS Prefix */
+		prefix.segbase=SegBase(ss);
+		prefix.mark|=PREFIX_SEG;
+		prefix.count++;
+		goto rep_again;
+	case 0x3e:			/* DS Prefix */
+		prefix.segbase=SegBase(ds);
+		prefix.mark|=PREFIX_SEG;
+		prefix.count++;
+		goto rep_again;
+	case 0x66:			/* Size Prefix */
+		prefix.count++;
+		prefix_66=!prefix_66;
+		goto rep_again;
+	case 0x6c:			/* REP INSB */
+		for (;(reg_cx && count>0);reg_cx--,count--) {
+			SaveMb(base_di+reg_di,IO_Read(reg_dx));
+			reg_di+=direct;
 		}
-	} while (again);
+		if (reg_cx && count<=0) goto countzero;
+		break;
+	case 0x6d:			/* REP INSW/D */
+		if (prefix_66) {
+			direct*=4;
+			for (;(reg_cx && count>0);reg_cx--,count--) {
+				SaveMb(base_di+reg_di+0,IO_Read(reg_dx+0));
+				SaveMb(base_di+reg_di+1,IO_Read(reg_dx+1));
+				SaveMb(base_di+reg_di+2,IO_Read(reg_dx+2));
+				SaveMb(base_di+reg_di+3,IO_Read(reg_dx+3));
+				reg_di+=direct;
+			}
+		} else {
+			direct*=2;
+			for (;(reg_cx && count>0);reg_cx--,count--) {
+				SaveMb(base_di+reg_di+0,IO_Read(reg_dx+0));
+				SaveMb(base_di+reg_di+1,IO_Read(reg_dx+1));
+				reg_di+=direct;
+			}
+		}
+		if (reg_cx && count<=0) goto countzero;
+		break;
+	case 0x6e:			/* REP OUTSB */
+		for (;(reg_cx && count>0);reg_cx--,count--) {
+			IO_Write(reg_dx,LoadMb(base_si+reg_si));
+			reg_si+=direct;
+		}	
+		if (reg_cx && count<=0) goto countzero;
+		break;
+	case 0x6f:			/* REP OUTSW/D */
+		if (prefix_66) {
+			direct*=4;
+			for (;(reg_cx && count>0);reg_cx--,count--) {
+				IO_Write(reg_dx+0,LoadMb(base_si+reg_si+0));
+				IO_Write(reg_dx+1,LoadMb(base_si+reg_si+1));
+				IO_Write(reg_dx+2,LoadMb(base_si+reg_si+2));
+				IO_Write(reg_dx+3,LoadMb(base_si+reg_si+3));
+				reg_si+=direct;
+			}
+		} else {
+			direct*=2;
+			for (;(reg_cx && count>0);reg_cx--,count--) {
+				IO_Write(reg_dx+0,LoadMb(base_si+reg_si+0));
+				IO_Write(reg_dx+1,LoadMb(base_si+reg_si+1));
+				reg_si+=direct;
+			}
+		}
+		if (reg_cx && count<=0) goto countzero;
+		break;
+	case 0xa4:			/* REP MOVSB */
+		for (;(reg_cx && count>0);reg_cx--,count--) {
+			SaveMb(base_di+reg_di,LoadMb(base_si+reg_si));
+			reg_si+=direct;reg_di+=direct;
+		}
+		if (reg_cx && count<=0) goto countzero;
+		break;
+	case 0xa5:			/* REP MOVSW/D */
+		if (prefix_66) {
+			direct*=4;
+			for (;(reg_cx && count>0);reg_cx--,count--) {
+				SaveMd(base_di+reg_di,LoadMd(base_si+reg_si));
+				reg_si+=direct;reg_di+=direct;
+			} 
+		} else {
+			direct*=2;
+			for (;(reg_cx && count>0);reg_cx--,count--) {
+				SaveMw(base_di+reg_di,LoadMw(base_si+reg_si));
+				reg_si+=direct;reg_di+=direct;
+			}
+		}
+		if (reg_cx && count<=0) goto countzero;
+		break;
+	case 0xa6:			/* REP CMPSB */
+		{	
+			if (!reg_cx) goto stopit;Bit8u op1,op2;
+			for (;(reg_cx && count>0);) {
+				op1=LoadMb(base_si+reg_si);op2=LoadMb(base_di+reg_di);
+				reg_cx--,count--;reg_si+=direct;reg_di+=direct;
+				if ((op1==op2)!=testz) break;
+			}
+			CMPB(op1,op2,LoadRb,0);
+			if ((op1==op2)!=testz) goto stopit;
+			if (reg_cx && count<=0) goto countzero;
+		}
+		break;
+	case 0xa7:			/* REP CMPSW */
+		{	
+			if (!reg_cx) goto stopit;
+			if (prefix_66) {
+				direct*=4;	
+				Bit32u op1,op2;
+				for (;(reg_cx && count>0);) {
+					op1=LoadMd(base_si+reg_si);op2=LoadMd(base_di+reg_di);
+					reg_cx--,count--;reg_si+=direct;reg_di+=direct;
+					if ((op1==op2)!=testz) break;
+				}
+				CMPD(op1,op2,LoadRd,0);
+				if ((op1==op2)!=testz) goto stopit;
+			} else {
+				direct*=2;
+				Bit16u op1,op2;
+				for (;(reg_cx && count>0);) {
+					op1=LoadMw(base_si+reg_si);op2=LoadMw(base_di+reg_di);
+					reg_cx--,count--;reg_si+=direct;reg_di+=direct;
+					if ((op1==op2)!=testz) break;
+				}
+				CMPW(op1,op2,LoadRw,0);
+				if ((op1==op2)!=testz) goto stopit;
+			}
+			if (reg_cx && count<=0) goto countzero;
+		}
+		break;
+	case 0xaa:			/* REP STOSB */
+		for (;(reg_cx && count>0);reg_cx--,count--) {
+			SaveMb(base_di+reg_di,reg_al);
+			reg_di+=direct;
+		}
+		if (reg_cx && count<=0) goto countzero;
+		break;
+	case 0xab:			/* REP STOSW */
+		if (prefix_66) {
+			direct*=4;	
+			for (;(reg_cx && count>0);reg_cx--,count--) {
+				SaveMd(base_di+reg_di,reg_eax);
+				reg_di+=direct;
+			}
+		} else {
+			direct*=2;	
+			for (;(reg_cx && count>0);reg_cx--,count--) {
+				SaveMw(base_di+reg_di,reg_ax);
+				reg_di+=direct;
+			}
+		}
+		if (reg_cx && count<=0) goto countzero;
+		break;
+	case 0xac:			/* REP LODSB */
+		for (;(reg_cx && count>0);reg_cx--,count--) {
+			reg_al=LoadMb(base_si+reg_si);
+			reg_si+=direct;
+		}	
+		if (reg_cx && count<=0) goto countzero;
+		break;
+	case 0xad:			/* REP LODSW */
+		if (prefix_66) {
+			direct*=4;
+			for (;(reg_cx && count>0);reg_cx--,count--) {
+				reg_eax=LoadMd(base_si+reg_si);
+				reg_si+=direct;
+			}	
+		} else {
+			direct*=2;
+			for (;(reg_cx && count>0);reg_cx--,count--) {
+				reg_ax=LoadMw(base_si+reg_si);
+				reg_si+=direct;
+			}	
+		}
+		if (reg_cx && count<=0) goto countzero;
+		break;
+	case 0xae:			/* REP SCASB */
+		{	
+			if (!reg_cx) goto stopit;Bit8u op2;
+			for (;(reg_cx && count>0);) {
+				op2=LoadMb(base_di+reg_di);
+				reg_cx--,count--;reg_di+=direct;
+				if ((reg_al==op2)!=testz) break;
+			}
+			CMPB(reg_al,op2,LoadRb,0);
+			if ((reg_al==op2)!=testz) goto stopit;
+			if (reg_cx && count<=0) goto countzero;
+		}
+		break;
+	case 0xaf:			/* REP SCASW */
+		{	
+			if (!reg_cx) goto stopit;
+			if (prefix_66) {
+				direct*=4;	
+				Bit32u op2;
+				for (;(reg_cx && count>0);) {
+					op2=LoadMd(base_di+reg_di);
+					reg_cx--,count--;reg_di+=direct;
+					if ((reg_eax==op2)!=testz) break;
+				}
+				CMPD(reg_eax,op2,LoadRd,0);
+				if ((reg_eax==op2)!=testz) goto stopit;
+			} else {
+				direct*=2;
+				Bit16u op2;
+				for (;(reg_cx && count>0);) {
+					op2=LoadMw(base_di+reg_di);
+					reg_cx--,count--;reg_di+=direct;
+					if ((reg_ax==op2)!=testz) break;
+				}
+				CMPW(reg_ax,op2,LoadRw,0);
+				if ((reg_ax==op2)!=testz) goto stopit;
+			}
+			if (reg_cx && count<=0) goto countzero;
+		}
+		break;
+	default:
+		IPPoint--;
+		LOG_DEBUG("Unhandled REP Prefix %X",Fetchb());
+
+	}
+stopit:
+	PrefixReset;
+	return count_remain-count;
+countzero:
+	IPPoint-=(prefix.count+2);		/* Rep instruction and whatever string instruction */
+	PrefixReset;
+	return count_remain;
 }
 
 //flags.io and nt shouldn't be compiled for 386 
+#ifdef CPU_386
 #define Save_Flagsw(FLAGW)											\
 {																	\
 	flags.type=t_UNKNOWN;											\
@@ -159,7 +376,25 @@ static INLINE void Rep_66(Bit16s direct,EAPoint from,EAPoint to) {
 		PIC_runIRQs();												\
 		LOADIP;														\
 	}																\
-	if (flags.tf) E_Exit("CPU:Trap Flag not supported");			\
+	if (flags.tf)	LOG_DEBUG("CPU:Trap Flag not supported");			\
 }
 
+#else 
 
+#define Save_Flagsw(FLAGW)											\
+{																	\
+	flags.type=t_UNKNOWN;											\
+	flags.cf	=(FLAGW & 0x001)>0;flags.pf	=(FLAGW & 0x004)>0;		\
+	flags.af	=(FLAGW & 0x010)>0;flags.zf	=(FLAGW & 0x040)>0;		\
+	flags.sf	=(FLAGW & 0x080)>0;flags.tf	=(FLAGW & 0x100)>0;		\
+	flags.intf	=(FLAGW & 0x200)>0;									\
+	flags.df	=(FLAGW & 0x400)>0;flags.of	=(FLAGW & 0x800)>0;		\
+	if (flags.intf && PIC_IRQCheck) {								\
+		SAVEIP;														\
+		PIC_runIRQs();												\
+		LOADIP;														\
+	}																\
+	if (flags.tf)	LOG_DEBUG("CPU:Trap Flag not supported");			\
+}
+
+#endif