Remember last prefix, for future SSE instruction decoding

include/cpu.h

@@ -358,6 +358,15 @@ struct TSS_32 {
     uint32_t ldt;                  /* The local descriptor table */
 } GCC_ATTRIBUTE(packed);
 
+/* Last prefix encountered, needed for Pentium III "mandatory opcode prefixes" needed to differentiate SSE instructions given the opcode.
+ * Keeping it small and sequential should help your C++ compiler optimize the switch statement you're probably going to use in the normal core code. */
+enum {
+	MP_NONE=0,
+	MP_66,
+	MP_F2,
+	MP_F3
+};
+
 #ifdef _MSC_VER
 #pragma pack()
 #endif

src/cpu/core_normal.cpp

@@ -100,6 +100,10 @@ Bitu cycle_count;
 	core.rep_zero=_ZERO;					\
 	goto restart_opcode;
 
+#define REMEMBER_PREFIX(_x) last_prefix = (_x)
+
+static uint8_t last_prefix;
+
 typedef PhysPt (*GetEAHandler)(void);
 
 static const uint32_t AddrMaskTable[2]={0x0000ffffu,0xffffffffu};
@@ -167,6 +171,7 @@ Bits CPU_Core_Normal_Run(void) {
 
 	while (CPU_Cycles-->0) {
 		LOADIP;
+		last_prefix=MP_NONE;
 		core.opcode_index=cpu.code.big*(Bitu)0x200u;
 		core.prefixes=cpu.code.big;
 		core.ea_table=&EATable[cpu.code.big*256u];

src/cpu/core_normal/prefix_none.h

@@ -337,6 +337,7 @@
         goto illegal_opcode;
 #endif
 #if CPU_CORE >= CPU_ARCHTYPE_386
+		REMEMBER_PREFIX(MP_66);
 		core.opcode_index=(cpu.code.big^0x1u)*0x200u;
 		goto restart_opcode;
 #endif
@@ -1282,9 +1283,11 @@
         goto opcode_f0;
 #endif
 	CASE_B(0xf2)												/* REPNZ */
+		REMEMBER_PREFIX(MP_F2);
 		DO_PREFIX_REP(false);	
 		break;		
 	CASE_B(0xf3)												/* REPZ */
+		REMEMBER_PREFIX(MP_F3);
 		DO_PREFIX_REP(true);	
 		break;		
 	CASE_B(0xf4)												/* HLT */

src/cpu/core_normal_286.cpp

@@ -119,6 +119,8 @@ extern Bitu cycle_count;
 	core.rep_zero=_ZERO;					\
 	goto restart_opcode;
 
+#define REMEMBER_PREFIX(_x)
+
 typedef PhysPt (*GetEAHandler)(void);
 
 static const uint32_t AddrMaskTable[2]={0x0000ffffu,0x0000ffffu};

src/cpu/core_normal_8086.cpp

@@ -149,6 +149,8 @@ extern Bitu cycle_count;
 	core.rep_zero=_ZERO;					\
 	goto restart_opcode;
 
+#define REMEMBER_PREFIX(_x)
+
 typedef PhysPt (*GetEAHandler)(void);
 
 static const uint32_t AddrMaskTable[2]={0x0000ffffu,0x0000ffffu};

src/cpu/core_prefetch.cpp

@@ -98,6 +98,10 @@ extern Bitu cycle_count;
 	core.rep_zero=_ZERO;					\
 	goto restart_opcode;
 
+#define REMEMBER_PREFIX(_x) last_prefix = (_x)
+
+static uint8_t last_prefix;
+
 typedef PhysPt (*GetEAHandler)(void);
 
 static const uint32_t AddrMaskTable[2]={0x0000ffffu,0xffffffffu};
@@ -203,6 +207,7 @@ Bits CPU_Core_Prefetch_Run(void) {
 			invalidate_pq=false;
 		}
 		LOADIP;
+		last_prefix=MP_NONE;
 		core.opcode_index=cpu.code.big*(Bitu)0x200u;
 		core.prefixes=cpu.code.big;
 		core.ea_table=&EATable[cpu.code.big*256u];

src/cpu/core_prefetch_286.cpp

@@ -122,6 +122,8 @@ Bits CPU_Core_Prefetch_Trap_Run(void);
 	core.rep_zero=_ZERO;					\
 	goto restart_opcode;
 
+#define REMEMBER_PREFIX(_x)
+
 typedef PhysPt (*GetEAHandler)(void);
 
 static const uint32_t AddrMaskTable[2]={0x0000ffffu,0x0000ffffu};

src/cpu/core_prefetch_8086.cpp

@@ -178,6 +178,8 @@ static struct {
 #define BaseDS		core.base_ds
 #define BaseSS		core.base_ss
 
+#define REMEMBER_PREFIX(_x)
+
 //#define PREFETCH_DEBUG
 
 #define MAX_PQ_SIZE 32

src/cpu/core_simple.cpp

@@ -85,6 +85,10 @@ extern Bitu cycle_count;
     core.rep_zero=_ZERO;                    \
     goto restart_opcode;
 
+#define REMEMBER_PREFIX(_x) last_prefix = (_x)
+
+static uint8_t last_prefix;
+
 typedef PhysPt (*GetEAHandler)(void);
 
 static const uint32_t AddrMaskTable[2]={0x0000ffffu,0xffffffffu};
@@ -176,6 +180,7 @@ Bits CPU_Core_Simple_Run(void) {
         /* Simple core optimizes for non-paged linear memory access and can break (segfault) if beyond end of memory */
         if (core.cseip >= safety_limit) break;
 
+        last_prefix=MP_NONE;
         core.opcode_index=cpu.code.big*0x200u;
         core.prefixes=cpu.code.big;
         core.ea_table=&EATable[cpu.code.big*256u];