simx86: handle RETl/RETlisp more efficiently.

Using O_POP1/O_POP2/O_POP2/O_POP3 in both rm and pm.
It's ok to pop into TheCPU.eip as that gets recomputed when
there is a fault anyway. Had to add RETlisp functionality
to O_POP2.

Author: bartoldeman

src/base/emu-i386/simx86/codegen-sim.c

@@ -2067,7 +2067,8 @@ static unsigned int Gen_sim(const IGen *IG, dosaddr_t mem_ref)
 		wkreg AR2;
 		unsigned int o = (mode & MNOREG) ? 0 : IG->p0;
 		long stackm = CPULONG(Ofs_STACKM);
-		GTRACE1("O_POP2",o);
+		int imm16 = (mode&MRETISP) ? IG->p0 : 0;
+		GTRACE2("O_POP2",o,imm16);
 		AR2.d = CPULONG(Ofs_XSS);
 #ifndef STACK_WRAP_MP	/* mask before incrementing */
 		SR1.d &= stackm;
@@ -2076,13 +2077,13 @@ static unsigned int Gen_sim(const IGen *IG, dosaddr_t mem_ref)
 			DR1.w.l = sim_read_word(AR2.d + SR1.d);
 			if (!(mode & MNOREG))
 				CPUWORD(o) = DR1.w.l;
-			SR1.d += 2;
+			SR1.d += 2 + imm16;
 		}
 		else {
 			DR1.d = sim_read_dword(AR2.d + SR1.d);
 			if (!(mode & MNOREG))
 				CPULONG(o) = DR1.d;
-			SR1.d += 4;
+			SR1.d += 4 + imm16;
 		}
 #ifdef STACK_WRAP_MP	/* mask after incrementing */
 		SR1.d &= stackm;

src/base/emu-i386/simx86/codegen-x86.c

@@ -1501,7 +1501,7 @@ arith0:		{
 			0x8d,0x14,0x0e,
 			// movw (%%edx,%%ebp,1),%%ax
 			0x66,0x8b,0x04,0x2a,
-/*0a*/			// leal 2(%%esi),%%esi
+			// leal 2(%%esi),%%esi
 			0x8d,0x76,0x02,
 #ifdef STACK_WRAP_MP	/* mask after incrementing */
 			// andl StackMask(%%ebx),%%esi
@@ -1517,7 +1517,7 @@ arith0:		{
 			0x8d,0x14,0x0e,
 			// movl (%%edx,%%ebp,1),%%eax
 			0x90,0x8b,0x04,0x2a,
-/*0a*/			// leal 4(%%esi),%%esi
+			// leal 4(%%esi),%%esi
 			0x8d,0x76,0x04,
 #ifdef STACK_WRAP_MP	/* mask after incrementing */
 			// andl StackMask(%%ebx),%%esi
@@ -1532,6 +1532,14 @@ arith0:		{
 		//	first do address calculation, then pop,
 		//	then store data, and last adjust stack
 		GNX(Cp, p, sz);
+		if (mode & MRETISP) {
+			// leal IG->p0(%%esi),%%esi
+			G2M(0x8d,0xb6,Cp); G4(IG->p0,Cp);
+#ifdef STACK_WRAP_MP	/* mask after incrementing */
+			// andl StackMask(%%ebx),%%esi
+			G3M(0x23,0x73,Ofs_STACKM,Cp);
+#endif
+		}
 		if (!(mode & MNOREG)) {
 			// mov{wl} %%{e}ax,offs(%%ebx)
 			Gen66(mode, Cp); G3M(0x89,0x43,IG->p0,Cp);

src/base/emu-i386/simx86/codegen.c

@@ -269,7 +269,7 @@ void Gen(int op, int mode, ...)
 		break;
 
 	case O_POP2:
-		if (!(mode & MNOREG))
+		if (!(mode & MNOREG) || (mode && MRETISP))
 			IG->p0 = va_arg(ap,unsigned int);
 		break;
 

src/base/emu-i386/simx86/interp.c

@@ -388,8 +388,6 @@ static unsigned int _JumpGen(unsigned int P2, int mode, int opc,
 		Gen(JLOOP_LINK, mode, opc, j_t, j_nt);
 		break;
 	case RETl: case RETlisp: // far ret, indirect
-		if (REALADDR()) AddrGen(A_SR_SH4, mode, Ofs_CS, Ofs_XCS);
-		/* fall through */
 	case JMPli: case CALLli: case INT: // far jmp/call, indirect
 		Gen(L_REG, mode, Ofs_EIP);
 		/* fall through */
@@ -1860,23 +1858,21 @@ intop3b:		{ int op = ArOpsFR[D_MO(opc)];
 /*c9*/	case LEAVE:
 			Gen(O_LEAVE, _mode); PC++;
 			break;
-/*ca*/	case RETlisp:	/* restartable */
-			if (!REALADDR()) {
-				/* pop from stack without adjusting esp */
-				Gen(O_POP1, _mode);
-				Gen(O_POP2, _mode|MNOREG);
-				Gen(O_POP2, _mode|MNOREG);
+/*ca*/	case RETlisp: {	/* restartable */
+			/* pop from stack without adjusting esp before A_SR_* */
+			int dr = (signed short)FetchW(PC+1);
+			Gen(O_POP1, _mode);
+			Gen(O_POP2, _mode, Ofs_EIP);
+			Gen(O_POP2, _mode|MNOREG|MRETISP, dr);
+			if (REALADDR())
+				AddrGen(A_SR_SH4, _mode, Ofs_CS, Ofs_XCS);
+			else
 				AddrGen(A_SR_PROT, _mode, Ofs_CS, P0);
-			}
-			{
-				int dr = (signed short)FetchW(PC+1);
-				Gen(O_POP, _mode);
-				Gen(S_REG, _mode, Ofs_EIP);
-				Gen(O_POP, _mode|MRETISP, dr);
-				PC = JumpGen(PC, _mode, opc, 3, P0, _flags);
-				if (debug_level('e')>2)
-					e_printf("RET_%d: ret=%08x\n",dr,TheCPU.eip);
-				if (TheCPU.err) return PC;
+			Gen(O_POP3, _mode);
+			PC = JumpGen(PC, _mode, opc, 3, P0, _flags);
+			if (debug_level('e')>2)
+				e_printf("RET_%d: ret=%08x\n",dr,TheCPU.eip);
+			if (TheCPU.err) return PC;
 			}
 			break;
 /*cc*/	case INT3:
@@ -1963,16 +1959,15 @@ intop3b:		{ int op = ArOpsFR[D_MO(opc)];
 		}
 
 /*cb*/	case RETl:
-			if (!REALADDR()) {
-			    /* pop from stack without adjusting esp */
-			    Gen(O_POP1, _mode);
-			    Gen(O_POP2, _mode|MNOREG);
-			    Gen(O_POP2, _mode|MNOREG);
-			    AddrGen(A_SR_PROT, _mode, Ofs_CS, P0);
-			}
-			Gen(O_POP, _mode);
-			Gen(S_REG, _mode, Ofs_EIP);
-			Gen(O_POP, _mode);
+			/* pop from stack without adjusting esp before A_SR_* */
+			Gen(O_POP1, _mode);
+			Gen(O_POP2, _mode, Ofs_EIP);
+			Gen(O_POP2, _mode|MNOREG);
+			if (REALADDR())
+				AddrGen(A_SR_SH4, _mode, Ofs_CS, Ofs_XCS);
+			else
+				AddrGen(A_SR_PROT, _mode, Ofs_CS, P0);
+			Gen(O_POP3, _mode);
 			PC = JumpGen(PC, _mode, opc, 1, P0, _flags);
 			if (debug_level('e')>1)
 			    e_printf("RET_FAR: ret=%04x:%08x\n",TheCPU.cs,TheCPU.eip);