diff --git a/src/386_common.h b/src/386_common.h index bbcfaa0..88a92fc 100644 --- a/src/386_common.h +++ b/src/386_common.h @@ -5,13 +5,13 @@ extern uint16_t ea_rseg; #define readmemb(s,a) ((readlookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF)?readmemb386l(s,a): *(uint8_t *)(readlookup2[(uint32_t)((s)+(a))>>12] + (uint32_t)((s) + (a))) ) -#define readmemq(s,a) ((readlookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a))&0xFFF)>0xFF8)?readmemql(s,a):*(uint64_t *)(readlookup2[(uint32_t)((s)+(a))>>12]+(uint32_t)((s)+(a)))) +#define readmemq(s,a) ((readlookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a)) & 7))?readmemql(s,a):*(uint64_t *)(readlookup2[(uint32_t)((s)+(a))>>12]+(uint32_t)((s)+(a)))) #define writememb(s,a,v) if (writelookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF) writememb386l(s,a,v); else *(uint8_t *)(writelookup2[(uint32_t)((s) + (a)) >> 12] + (uint32_t)((s) + (a))) = v -#define writememw(s,a,v) if (writelookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a))&0xFFF)>0xFFE) writememwl(s,a,v); else *(uint16_t *)(writelookup2[(uint32_t)((s) + (a)) >> 12] + (uint32_t)((s) + (a))) = v -#define writememl(s,a,v) if (writelookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a))&0xFFF)>0xFFC) writememll(s,a,v); else *(uint32_t *)(writelookup2[(uint32_t)((s) + (a)) >> 12] + (uint32_t)((s) + (a))) = v -#define writememq(s,a,v) if (writelookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a))&0xFFF)>0xFF8) writememql(s,a,v); else *(uint64_t *)(writelookup2[(uint32_t)((s) + (a)) >> 12] + (uint32_t)((s) + (a))) = v +#define writememw(s,a,v) if (writelookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a)) & 1)) writememwl(s,a,v); else *(uint16_t *)(writelookup2[(uint32_t)((s) + (a)) >> 12] + (uint32_t)((s) + (a))) = v +#define writememl(s,a,v) if (writelookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a)) & 3)) writememll(s,a,v); else *(uint32_t *)(writelookup2[(uint32_t)((s) + (a)) >> 12] + (uint32_t)((s) + (a))) = v +#define writememq(s,a,v) if (writelookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a)) & 7)) writememql(s,a,v); else *(uint64_t *)(writelookup2[(uint32_t)((s) + (a)) >> 12] + (uint32_t)((s) + (a))) = v int checkio(int port); @@ -88,8 +88,8 @@ static inline uint16_t fastreadw(uint32_t a) uint16_t val; if ((a&0xFFF)>0xFFE) { - val = readmemb(0, a); - val |= (readmemb(0, a + 1) << 8); + val = fastreadb(a); + val |= (fastreadb(a + 1) << 8); return val; } if ((a>>12)==pccache) return *((uint16_t *)&pccache2[a]); @@ -119,10 +119,8 @@ static inline uint32_t fastreadl(uint32_t a) } return *((uint32_t *)&pccache2[a]); } - val =readmemb(0,a); - val |=(readmemb(0,a+1)<<8); - val |=(readmemb(0,a+2)<<16); - val |=(readmemb(0,a+3)<<24); + val = fastreadw(a); + val |= (fastreadw(a + 2) << 16); return val; } diff --git a/src/codegen_ops_x86-64.h b/src/codegen_ops_x86-64.h index 3742053..e0cb874 100644 --- a/src/codegen_ops_x86-64.h +++ b/src/codegen_ops_x86-64.h @@ -1069,16 +1069,14 @@ static inline void MEM_LOAD_ADDR_EA_W(x86seg *seg) addbyte(0x8d); addbyte(0x34); addbyte(0x08); - addbyte(0x67); /*LEA EDI, 1[ESI]*/ - addbyte(0x8d); - addbyte(0x7e); - addbyte(0x01); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xfff*/ + addbyte(0xf7); /*TEST EDI, 1*/ addbyte(0xc7); - addlong(0xfff); + addlong(1); if (IS_32_ADDR(readlookup2)) { addbyte(0x67); /*MOV RSI, readlookup2[ESI*8]*/ @@ -1098,18 +1096,17 @@ static inline void MEM_LOAD_ADDR_EA_W(x86seg *seg) addbyte(0x34); addbyte(0xf2); } - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+5+2); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+4+2); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(5+2); - addbyte(0x66); /*MOV AX,-1[RDI+RSI]*/ + addbyte(4+2); + addbyte(0x66); /*MOV AX,[RDI+RSI]*/ addbyte(0x8b); - addbyte(0x44); + addbyte(0x04); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-1); addbyte(0xeb); /*JMP done*/ addbyte(2+2+12+4+6); /*slowpath:*/ @@ -1158,16 +1155,14 @@ static inline void MEM_LOAD_ADDR_EA_L(x86seg *seg) addbyte(0x8d); addbyte(0x34); addbyte(0x08); - addbyte(0x67); /*LEA EDI, 3[ESI]*/ - addbyte(0x8d); - addbyte(0x7e); - addbyte(0x03); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xffc*/ + addbyte(0xf7); /*TEST EDI, 3*/ addbyte(0xc7); - addlong(0xffc); + addlong(3); if (IS_32_ADDR(readlookup2)) { addbyte(0x67); /*MOV RSI, readlookup2[ESI*8]*/ @@ -1187,17 +1182,16 @@ static inline void MEM_LOAD_ADDR_EA_L(x86seg *seg) addbyte(0x34); addbyte(0xf2); } - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+4+2); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+3+2); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(4+2); - addbyte(0x8b); /*MOV EAX,-3[RDI+RSI]*/ - addbyte(0x44); + addbyte(3+2); + addbyte(0x8b); /*MOV EAX,[RDI+RSI]*/ + addbyte(0x04); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-3); addbyte(0xeb); /*JMP done*/ addbyte(2+2+12+4+6); /*slowpath:*/ @@ -1239,16 +1233,14 @@ static inline void MEM_LOAD_ADDR_EA_Q(x86seg *seg) addbyte(0x8d); addbyte(0x34); addbyte(0x08); - addbyte(0x67); /*LEA EDI, 7[ESI]*/ - addbyte(0x8d); - addbyte(0x7e); - addbyte(0x07); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xff8*/ + addbyte(0xf7); /*TEST EDI, 7*/ addbyte(0xc7); - addlong(0xff8); + addlong(7); if (IS_32_ADDR(readlookup2)) { addbyte(0x67); /*MOV RSI, readlookup2[ESI*8]*/ @@ -1268,18 +1260,17 @@ static inline void MEM_LOAD_ADDR_EA_Q(x86seg *seg) addbyte(0x34); addbyte(0xf2); } - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+5+2); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+4+2); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(5+2); - addbyte(0x48); /*MOV RAX,-7[RDI+RSI]*/ + addbyte(4+2); + addbyte(0x48); /*MOV RAX,[RDI+RSI]*/ addbyte(0x8b); - addbyte(0x44); + addbyte(0x04); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-7); addbyte(0xeb); /*JMP done*/ addbyte(2+2+12+4+6); /*slowpath:*/ @@ -1447,16 +1438,14 @@ static inline void MEM_STORE_ADDR_EA_W(x86seg *seg, int host_reg) addbyte(0x8d); addbyte(0x34); addbyte(0x08); - addbyte(0x67); /*LEA EDI, 1[ESI]*/ - addbyte(0x8d); - addbyte(0x7e); - addbyte(0x01); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xfff*/ + addbyte(0xf7); /*TEST EDI, 1*/ addbyte(0xc7); - addlong(0xfff); + addlong(1); if (IS_32_ADDR(writelookup2)) { addbyte(0x67); /*MOV RSI, writelookup2[ESI*8]*/ @@ -1476,29 +1465,27 @@ static inline void MEM_STORE_ADDR_EA_W(x86seg *seg, int host_reg) addbyte(0x34); addbyte(0xf2); } - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+((host_reg & 8) ? 6:5)+2); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+((host_reg & 8) ? 5:4)+2); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(((host_reg & 8) ? 6:5)+2); + addbyte(((host_reg & 8) ? 5:4)+2); if (host_reg & 8) { - addbyte(0x66); /*MOV -1[RDI+RSI],host_reg*/ + addbyte(0x66); /*MOV [RDI+RSI],host_reg*/ addbyte(0x44); addbyte(0x89); - addbyte(0x44 | ((host_reg & 7) << 3)); + addbyte(0x04 | ((host_reg & 7) << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-1); } else { - addbyte(0x66); /*MOV -1[RDI+RSI],host_reg*/ + addbyte(0x66); /*MOV [RDI+RSI],host_reg*/ addbyte(0x89); - addbyte(0x44 | (host_reg << 3)); + addbyte(0x04 | (host_reg << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-1); } addbyte(0xeb); /*JMP done*/ addbyte(2+2+3+12+4+6); @@ -1542,16 +1529,14 @@ static inline void MEM_STORE_ADDR_EA_L(x86seg *seg, int host_reg) addbyte(0x8d); addbyte(0x34); addbyte(0x08); - addbyte(0x67); /*LEA EDI, 3[ESI]*/ - addbyte(0x8d); - addbyte(0x7e); - addbyte(0x03); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xffc*/ + addbyte(0xf7); /*TEST EDI, 3*/ addbyte(0xc7); - addlong(0xffc); + addlong(3); if (IS_32_ADDR(writelookup2)) { addbyte(0x67); /*MOV RSI, writelookup2[ESI*8]*/ @@ -1571,27 +1556,25 @@ static inline void MEM_STORE_ADDR_EA_L(x86seg *seg, int host_reg) addbyte(0x34); addbyte(0xf2); } - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+((host_reg & 8) ? 5:4)+2); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+((host_reg & 8) ? 4:3)+2); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(((host_reg & 8) ? 5:4)+2); + addbyte(((host_reg & 8) ? 4:3)+2); if (host_reg & 8) { addbyte(0x44); /*MOV -3[RDI+RSI],host_reg*/ addbyte(0x89); - addbyte(0x44 | ((host_reg & 7) << 3)); + addbyte(0x04 | ((host_reg & 7) << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-3); } else { addbyte(0x89); /*MOV -3[RDI+RSI],host_reg*/ - addbyte(0x44 | (host_reg << 3)); + addbyte(0x04 | (host_reg << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-3); } addbyte(0xeb); /*JMP done*/ addbyte(2+2+3+12+4+6); @@ -1635,16 +1618,14 @@ static inline void MEM_STORE_ADDR_EA_Q(x86seg *seg, int host_reg, int host_reg2) addbyte(0x8d); addbyte(0x34); addbyte(0x08); - addbyte(0x67); /*LEA EDI, 7[ESI]*/ - addbyte(0x8d); - addbyte(0x7e); - addbyte(0x07); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xff8*/ + addbyte(0xf7); /*TEST EDI, 7*/ addbyte(0xc7); - addlong(0xff8); + addlong(7); if (IS_32_ADDR(writelookup2)) { addbyte(0x67); /*MOV RSI, writelookup2[ESI*8]*/ @@ -1664,28 +1645,26 @@ static inline void MEM_STORE_ADDR_EA_Q(x86seg *seg, int host_reg, int host_reg2) addbyte(0x34); addbyte(0xf2); } - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+5+2); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+4+2); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(5+2); + addbyte(4+2); if (host_reg & 8) { - addbyte(0x4c); /*MOV -7[RDI+RSI],host_reg*/ + addbyte(0x4c); /*MOV [RDI+RSI],host_reg*/ addbyte(0x89); - addbyte(0x44 | ((host_reg & 7) << 3)); + addbyte(0x04 | ((host_reg & 7) << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-7); } else { - addbyte(0x48); /*MOV -3[RDI+RSI],host_reg*/ + addbyte(0x48); /*MOV [RDI+RSI],host_reg*/ addbyte(0x89); - addbyte(0x44 | (host_reg << 3)); + addbyte(0x04 | (host_reg << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-7); } addbyte(0xeb); /*JMP done*/ addbyte(2+2+3+12+4+6); @@ -5823,16 +5802,14 @@ static inline int MEM_LOAD_ADDR_EA_W_NO_ABRT(x86seg *seg) addbyte(0x8d); addbyte(0x34); addbyte(0x08); - addbyte(0x67); /*LEA EDI, 1[ESI]*/ - addbyte(0x8d); - addbyte(0x7e); - addbyte(0x01); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xfff*/ + addbyte(0xf7); /*TEST EDI, 1*/ addbyte(0xc7); - addlong(0xfff); + addlong(1); if (IS_32_ADDR(readlookup2)) { addbyte(0x67); /*MOV RSI, readlookup2[ESI*8]*/ @@ -5852,18 +5829,17 @@ static inline int MEM_LOAD_ADDR_EA_W_NO_ABRT(x86seg *seg) addbyte(0x34); addbyte(0xf2); } - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+5+2); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+4+2); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(5+2); - addbyte(0x66); /*MOV AX,-1[RDI+RSI]*/ + addbyte(4+2); + addbyte(0x66); /*MOV AX,[RDI+RSI]*/ addbyte(0x8b); - addbyte(0x44); + addbyte(0x04); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-1); addbyte(0xeb); /*JMP done*/ addbyte(2+2+12); /*slowpath:*/ @@ -5904,16 +5880,14 @@ static inline int MEM_LOAD_ADDR_EA_L_NO_ABRT(x86seg *seg) addbyte(0x8d); addbyte(0x34); addbyte(0x08); - addbyte(0x67); /*LEA EDI, 3[ESI]*/ - addbyte(0x8d); - addbyte(0x7e); - addbyte(0x03); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xffc*/ + addbyte(0xf7); /*TEST EDI, 1*/ addbyte(0xc7); - addlong(0xffc); + addlong(1); if (IS_32_ADDR(readlookup2)) { addbyte(0x67); /*MOV RSI, readlookup2[ESI*8]*/ @@ -5933,17 +5907,16 @@ static inline int MEM_LOAD_ADDR_EA_L_NO_ABRT(x86seg *seg) addbyte(0x34); addbyte(0xf2); } - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+4+2); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+3+2); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(4+2); - addbyte(0x8b); /*MOV EAX,-3[RDI+RSI]*/ - addbyte(0x44); + addbyte(3+2); + addbyte(0x8b); /*MOV EAX,[RDI+RSI]*/ + addbyte(0x04); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-3); addbyte(0xeb); /*JMP done*/ addbyte(2+2+12); /*slowpath:*/ @@ -6084,16 +6057,14 @@ static inline void MEM_STORE_ADDR_EA_W_NO_ABRT(x86seg *seg, int host_reg) addbyte(0x8d); addbyte(0x34); addbyte(0x18); - addbyte(0x67); /*LEA EDI, 1[ESI]*/ - addbyte(0x8d); - addbyte(0x7e); - addbyte(0x01); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xfff*/ + addbyte(0xf7); /*TEST EDI, 1*/ addbyte(0xc7); - addlong(0xfff); + addlong(1); if (IS_32_ADDR(writelookup2)) { addbyte(0x67); /*MOV RSI, writelookup2[ESI*8]*/ @@ -6113,29 +6084,27 @@ static inline void MEM_STORE_ADDR_EA_W_NO_ABRT(x86seg *seg, int host_reg) addbyte(0x34); addbyte(0xf2); } - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+((host_reg & 8) ? 6:5)+2); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+((host_reg & 8) ? 5:4)+2); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(((host_reg & 8) ? 6:5)+2); + addbyte(((host_reg & 8) ? 5:4)+2); if (host_reg & 8) { - addbyte(0x66); /*MOV -1[RDI+RSI],host_reg*/ + addbyte(0x66); /*MOV [RDI+RSI],host_reg*/ addbyte(0x44); addbyte(0x89); - addbyte(0x44 | ((host_reg & 7) << 3)); + addbyte(0x04 | ((host_reg & 7) << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-1); } else { - addbyte(0x66); /*MOV -1[RDI+RSI],host_reg*/ + addbyte(0x66); /*MOV [RDI+RSI],host_reg*/ addbyte(0x89); - addbyte(0x44 | (host_reg << 3)); + addbyte(0x04 | (host_reg << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-1); } addbyte(0xeb); /*JMP done*/ addbyte(2+2+3+12); @@ -6172,16 +6141,14 @@ static inline void MEM_STORE_ADDR_EA_L_NO_ABRT(x86seg *seg, int host_reg) addbyte(0x8d); addbyte(0x34); addbyte(0x18); - addbyte(0x67); /*LEA EDI, 3[ESI]*/ - addbyte(0x8d); - addbyte(0x7e); - addbyte(0x03); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xffc*/ + addbyte(0xf7); /*TEST EDI, 3*/ addbyte(0xc7); - addlong(0xffc); + addlong(3); if (IS_32_ADDR(writelookup2)) { addbyte(0x67); /*MOV RSI, writelookup2[ESI*8]*/ @@ -6201,27 +6168,25 @@ static inline void MEM_STORE_ADDR_EA_L_NO_ABRT(x86seg *seg, int host_reg) addbyte(0x34); addbyte(0xf2); } - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+((host_reg & 8) ? 5:4)+2); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+((host_reg & 8) ? 4:3)+2); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(((host_reg & 8) ? 5:4)+2); + addbyte(((host_reg & 8) ? 4:3)+2); if (host_reg & 8) { - addbyte(0x44); /*MOV -3[RDI+RSI],host_reg*/ + addbyte(0x44); /*MOV [RDI+RSI],host_reg*/ addbyte(0x89); - addbyte(0x44 | ((host_reg & 7) << 3)); + addbyte(0x04 | ((host_reg & 7) << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-3); } else { - addbyte(0x89); /*MOV -3[RDI+RSI],host_reg*/ - addbyte(0x44 | (host_reg << 3)); + addbyte(0x89); /*MOV [RDI+RSI],host_reg*/ + addbyte(0x04 | (host_reg << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-3); } addbyte(0xeb); /*JMP done*/ addbyte(2+2+3+12); diff --git a/src/codegen_x86.c b/src/codegen_x86.c index 841ef7e..9739da8 100644 --- a/src/codegen_x86.c +++ b/src/codegen_x86.c @@ -137,31 +137,29 @@ static uint32_t gen_MEM_LOAD_ADDR_EA_W() addbyte(0xd6); addbyte(0x01); /*ADDL EDX, EAX*/ addbyte(0xc2); - addbyte(0x8d); /*LEA EDI, 1[EDX]*/ - addbyte(0x7a); - addbyte(0x01); + addbyte(0x89); /*MOV EDI, EDX*/ + addbyte(0xd7); addbyte(0xc1); /*SHR EDX, 12*/ addbyte(0xea); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xfff*/ + addbyte(0xf7); /*TEST EDI, 1*/ addbyte(0xc7); - addlong(0xfff); + addlong(1); addbyte(0x8b); /*MOV EDX, readlookup2[EDX*4]*/ addbyte(0x14); addbyte(0x95); addlong((uint32_t)readlookup2); - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+5+1); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+4+1); addbyte(0x83); /*CMP EDX, -1*/ addbyte(0xfa); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(5+1); - addbyte(0x0f); /*MOVZX EAX, -1[EDX+EDI]W*/ + addbyte(4+1); + addbyte(0x0f); /*MOVZX EAX, [EDX+EDI]W*/ addbyte(0xb7); - addbyte(0x44); + addbyte(0x04); addbyte(0x3a); - addbyte(-1); addbyte(0xc3); /*RET*/ addbyte(0x50); /*slowpath: PUSH EAX*/ @@ -194,30 +192,28 @@ static uint32_t gen_MEM_LOAD_ADDR_EA_L() addbyte(0xd6); addbyte(0x01); /*ADDL EDX, EAX*/ addbyte(0xc2); - addbyte(0x8d); /*LEA EDI, 3[EDX]*/ - addbyte(0x7a); - addbyte(0x03); + addbyte(0x89); /*MOV EDI, EDX*/ + addbyte(0xd7); addbyte(0xc1); /*SHR EDX, 12*/ addbyte(0xea); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xffc*/ + addbyte(0xf7); /*TEST EDI, 3*/ addbyte(0xc7); - addlong(0xffc); + addlong(3); addbyte(0x8b); /*MOV EDX, readlookup2[EDX*4]*/ addbyte(0x14); addbyte(0x95); addlong((uint32_t)readlookup2); - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+4+1); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+3+1); addbyte(0x83); /*CMP EDX, -1*/ addbyte(0xfa); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(4+1); - addbyte(0x8b); /*MOV EAX, -3[EDX+EDI]*/ - addbyte(0x44); + addbyte(3+1); + addbyte(0x8b); /*MOV EAX, [EDX+EDI]*/ + addbyte(0x04); addbyte(0x3a); - addbyte(-3); addbyte(0xc3); /*RET*/ addbyte(0x50); /*slowpath: PUSH EAX*/ @@ -247,34 +243,32 @@ static uint32_t gen_MEM_LOAD_ADDR_EA_Q() addbyte(0xd6); addbyte(0x01); /*ADDL EDX, EAX*/ addbyte(0xc2); - addbyte(0x8d); /*LEA EDI, 7[EDX]*/ - addbyte(0x7a); - addbyte(0x07); + addbyte(0x89); /*MOV EDI, EDX*/ + addbyte(0xd7); addbyte(0xc1); /*SHR EDX, 12*/ addbyte(0xea); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xff8*/ + addbyte(0xf7); /*TEST EDI, 7*/ addbyte(0xc7); - addlong(0xff8); + addlong(7); addbyte(0x8b); /*MOV EDX, readlookup2[EDX*4]*/ addbyte(0x14); addbyte(0x95); addlong((uint32_t)readlookup2); - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+4+4+1); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+3+4+1); addbyte(0x83); /*CMP EDX, -1*/ addbyte(0xfa); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(4+4+1); + addbyte(3+4+1); addbyte(0x8b); /*MOV EAX, [EDX+EDI]*/ - addbyte(0x44); + addbyte(0x04); addbyte(0x3a); - addbyte(-7); addbyte(0x8b); /*MOV EDX, [EDX+EDI+4]*/ addbyte(0x54); addbyte(0x3a); - addbyte(-7+4); + addbyte(4); addbyte(0xc3); /*RET*/ addbyte(0x50); /*slowpath: PUSH EAX*/ @@ -353,31 +347,29 @@ static uint32_t gen_MEM_STORE_ADDR_EA_W() addbyte(0xf3); addbyte(0x01); /*ADDL ESI, EAX*/ addbyte(0xc0 | (REG_EAX << 3) | REG_ESI); - addbyte(0x8d); /*LEA EDI, 1[ESI]*/ - addbyte(0x7e); - addbyte(0x01); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xfff*/ + addbyte(0xf7); /*TEST EDI, 1*/ addbyte(0xc7); - addlong(0xfff); + addlong(1); addbyte(0x8b); /*MOV ESI, readlookup2[ESI*4]*/ addbyte(0x04 | (REG_ESI << 3)); addbyte(0x85 | (REG_ESI << 3)); addlong((uint32_t)writelookup2); - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+5+1); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+4+1); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(5+1); - addbyte(0x66); /*MOV -1[EDI+ESI],CX*/ + addbyte(4+1); + addbyte(0x66); /*MOV [EDI+ESI],CX*/ addbyte(0x89); - addbyte(0x44 | (REG_CX << 3)); + addbyte(0x04 | (REG_CX << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-1); addbyte(0xc3); /*RET*/ addbyte(0x51); /*slowpath: PUSH ECX*/ @@ -409,30 +401,28 @@ static uint32_t gen_MEM_STORE_ADDR_EA_L() addbyte(0xf3); addbyte(0x01); /*ADDL ESI, EAX*/ addbyte(0xc0 | (REG_EAX << 3) | REG_ESI); - addbyte(0x8d); /*LEA EDI, 3[ESI]*/ - addbyte(0x7e); - addbyte(0x03); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xffc*/ + addbyte(0xf7); /*TEST EDI, 3*/ addbyte(0xc7); - addlong(0xffc); + addlong(3); addbyte(0x8b); /*MOV ESI, readlookup2[ESI*4]*/ addbyte(0x04 | (REG_ESI << 3)); addbyte(0x85 | (REG_ESI << 3)); addlong((uint32_t)writelookup2); - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+4+1); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+3+1); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(4+1); - addbyte(0x89); /*MOV -3[EDI+ESI],ECX*/ - addbyte(0x44 | (REG_ECX << 3)); + addbyte(3+1); + addbyte(0x89); /*MOV [EDI+ESI],ECX*/ + addbyte(0x04 | (REG_ECX << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-3); addbyte(0xc3); /*RET*/ addbyte(0x51); /*slowpath: PUSH ECX*/ @@ -464,34 +454,32 @@ static uint32_t gen_MEM_STORE_ADDR_EA_Q() addbyte(0xf2); addbyte(0x01); /*ADDL ESI, EAX*/ addbyte(0xc0 | (REG_EAX << 3) | REG_ESI); - addbyte(0x8d); /*LEA EDI, 7[ESI]*/ - addbyte(0x7e); - addbyte(0x07); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xff8*/ + addbyte(0xf7); /*TEST EDI, 7*/ addbyte(0xc7); - addlong(0xff8); + addlong(7); addbyte(0x8b); /*MOV ESI, readlookup2[ESI*4]*/ addbyte(0x04 | (REG_ESI << 3)); addbyte(0x85 | (REG_ESI << 3)); addlong((uint32_t)writelookup2); - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+4+4+1); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+3+4+1); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(4+4+1); - addbyte(0x89); /*MOV -7[EDI+ESI],EBX*/ - addbyte(0x44 | (REG_EBX << 3)); + addbyte(3+4+1); + addbyte(0x89); /*MOV [EDI+ESI],EBX*/ + addbyte(0x04 | (REG_EBX << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-7); - addbyte(0x89); /*MOV -7[EDI+ESI],EBX*/ + addbyte(0x89); /*MOV 4[EDI+ESI],EBX*/ addbyte(0x44 | (REG_ECX << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-7+4); + addbyte(4); addbyte(0xc3); /*RET*/ addbyte(0x51); /*slowpath: PUSH ECX*/ @@ -515,7 +503,9 @@ static uint32_t gen_MEM_STORE_ADDR_EA_Q() return addr; } +#ifndef RELEASE_BUILD static char gen_MEM_LOAD_ADDR_EA_B_NO_ABRT_err[] = "gen_MEM_LOAD_ADDR_EA_B_NO_ABRT aborted\n"; +#endif static uint32_t gen_MEM_LOAD_ADDR_EA_B_NO_ABRT() { uint32_t addr = (uint32_t)&codeblock[block_current].data[block_pos]; @@ -577,7 +567,9 @@ static uint32_t gen_MEM_LOAD_ADDR_EA_B_NO_ABRT() return addr; } +#ifndef RELEASE_BUILD static char gen_MEM_LOAD_ADDR_EA_W_NO_ABRT_err[] = "gen_MEM_LOAD_ADDR_EA_W_NO_ABRT aborted\n"; +#endif static uint32_t gen_MEM_LOAD_ADDR_EA_W_NO_ABRT() { uint32_t addr = (uint32_t)&codeblock[block_current].data[block_pos]; @@ -586,31 +578,29 @@ static uint32_t gen_MEM_LOAD_ADDR_EA_W_NO_ABRT() addbyte(0xd6); addbyte(0x01); /*ADDL EDX, EAX*/ addbyte(0xc2); - addbyte(0x8d); /*LEA EDI, 1[EDX]*/ - addbyte(0x7a); - addbyte(0x01); + addbyte(0x89); /*MOV EDI, EDX*/ + addbyte(0xd7); addbyte(0xc1); /*SHR EDX, 12*/ addbyte(0xea); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xfff*/ + addbyte(0xf7); /*TEST EDI, 1*/ addbyte(0xc7); - addlong(0xfff); + addlong(1); addbyte(0x8b); /*MOV EDX, readlookup2[EDX*4]*/ addbyte(0x14); addbyte(0x95); addlong((uint32_t)readlookup2); - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+5+1); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+4+1); addbyte(0x83); /*CMP EDX, -1*/ addbyte(0xfa); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(5+1); - addbyte(0x0f); /*MOVZX EEX, -1[EDX+EDI]W*/ + addbyte(4+1); + addbyte(0x0f); /*MOVZX ECX, [EDX+EDI]W*/ addbyte(0xb7); - addbyte(0x4c); + addbyte(0x0c); addbyte(0x3a); - addbyte(-1); addbyte(0xc3); /*RET*/ addbyte(0x50); /*slowpath: PUSH EAX*/ @@ -646,7 +636,9 @@ static uint32_t gen_MEM_LOAD_ADDR_EA_W_NO_ABRT() return addr; } +#ifndef RELEASE_BUILD static char gen_MEM_LOAD_ADDR_EA_L_NO_ABRT_err[] = "gen_MEM_LOAD_ADDR_EA_L_NO_ABRT aborted\n"; +#endif static uint32_t gen_MEM_LOAD_ADDR_EA_L_NO_ABRT() { uint32_t addr = (uint32_t)&codeblock[block_current].data[block_pos]; @@ -655,30 +647,28 @@ static uint32_t gen_MEM_LOAD_ADDR_EA_L_NO_ABRT() addbyte(0xd6); addbyte(0x01); /*ADDL EDX, EAX*/ addbyte(0xc2); - addbyte(0x8d); /*LEA EDI, 3[EDX]*/ - addbyte(0x7a); - addbyte(0x03); + addbyte(0x89); /*MOV EDI, EDX*/ + addbyte(0xd7); addbyte(0xc1); /*SHR EDX, 12*/ addbyte(0xea); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xffc*/ + addbyte(0xf7); /*TEST EDI, 3*/ addbyte(0xc7); - addlong(0xffc); + addlong(3); addbyte(0x8b); /*MOV EDX, readlookup2[EDX*4]*/ addbyte(0x14); addbyte(0x95); addlong((uint32_t)readlookup2); - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+4+1); + addbyte(0x75); /*JE slowpath*/ + addbyte(3+2+3+1); addbyte(0x83); /*CMP EDX, -1*/ addbyte(0xfa); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(4+1); - addbyte(0x8b); /*MOV ECX, -3[EDX+EDI]*/ - addbyte(0x4c); + addbyte(3+1); + addbyte(0x8b); /*MOV ECX, [EDX+EDI]*/ + addbyte(0x0c); addbyte(0x3a); - addbyte(-3); addbyte(0xc3); /*RET*/ addbyte(0x50); /*slowpath: PUSH EAX*/ @@ -714,7 +704,9 @@ static uint32_t gen_MEM_LOAD_ADDR_EA_L_NO_ABRT() return addr; } +#ifndef RELEASE_BUILD static char gen_MEM_STORE_ADDR_EA_B_NO_ABRT_err[] = "gen_MEM_STORE_ADDR_EA_B_NO_ABRT aborted\n"; +#endif static uint32_t gen_MEM_STORE_ADDR_EA_B_NO_ABRT() { uint32_t addr = (uint32_t)&codeblock[block_current].data[block_pos]; @@ -772,7 +764,9 @@ static uint32_t gen_MEM_STORE_ADDR_EA_B_NO_ABRT() return addr; } +#ifndef RELEASE_BUILD static char gen_MEM_STORE_ADDR_EA_W_NO_ABRT_err[] = "gen_MEM_STORE_ADDR_EA_W_NO_ABRT aborted\n"; +#endif static uint32_t gen_MEM_STORE_ADDR_EA_W_NO_ABRT() { uint32_t addr = (uint32_t)&codeblock[block_current].data[block_pos]; @@ -782,31 +776,29 @@ static uint32_t gen_MEM_STORE_ADDR_EA_W_NO_ABRT() addbyte(0xf3); addbyte(0x01); /*ADDL ESI, EAX*/ addbyte(0xc0 | (REG_EAX << 3) | REG_ESI); - addbyte(0x8d); /*LEA EDI, 1[ESI]*/ - addbyte(0x7e); - addbyte(0x01); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xfff*/ + addbyte(0xf7); /*TEST EDI, 1*/ addbyte(0xc7); - addlong(0xfff); + addlong(1); addbyte(0x8b); /*MOV ESI, readlookup2[ESI*4]*/ addbyte(0x04 | (REG_ESI << 3)); addbyte(0x85 | (REG_ESI << 3)); addlong((uint32_t)writelookup2); - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+5+1); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+4+1); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(5+1); - addbyte(0x66); /*MOV -1[EDI+ESI],CX*/ + addbyte(4+1); + addbyte(0x66); /*MOV [EDI+ESI],CX*/ addbyte(0x89); - addbyte(0x44 | (REG_CX << 3)); + addbyte(0x04 | (REG_CX << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-1); addbyte(0xc3); /*RET*/ addbyte(0x51); /*slowpath: PUSH ECX*/ @@ -838,7 +830,9 @@ static uint32_t gen_MEM_STORE_ADDR_EA_W_NO_ABRT() return addr; } +#ifndef RELEASE_BUILD static char gen_MEM_STORE_ADDR_EA_L_NO_ABRT_err[] = "gen_MEM_STORE_ADDR_EA_L_NO_ABRT aborted\n"; +#endif static uint32_t gen_MEM_STORE_ADDR_EA_L_NO_ABRT() { uint32_t addr = (uint32_t)&codeblock[block_current].data[block_pos]; @@ -848,30 +842,28 @@ static uint32_t gen_MEM_STORE_ADDR_EA_L_NO_ABRT() addbyte(0xf3); addbyte(0x01); /*ADDL ESI, EAX*/ addbyte(0xc0 | (REG_EAX << 3) | REG_ESI); - addbyte(0x8d); /*LEA EDI, 3[ESI]*/ - addbyte(0x7e); - addbyte(0x03); + addbyte(0x89); /*MOV EDI, ESI*/ + addbyte(0xf7); addbyte(0xc1); /*SHR ESI, 12*/ addbyte(0xe8 | REG_ESI); addbyte(12); - addbyte(0xf7); /*TEST EDI, 0xffc*/ + addbyte(0xf7); /*TEST EDI, 3*/ addbyte(0xc7); - addlong(0xffc); + addlong(3); addbyte(0x8b); /*MOV ESI, readlookup2[ESI*4]*/ addbyte(0x04 | (REG_ESI << 3)); addbyte(0x85 | (REG_ESI << 3)); addlong((uint32_t)writelookup2); - addbyte(0x74); /*JE slowpath*/ - addbyte(3+2+4+1); + addbyte(0x75); /*JNE slowpath*/ + addbyte(3+2+3+1); addbyte(0x83); /*CMP ESI, -1*/ addbyte(0xf8 | REG_ESI); addbyte(-1); addbyte(0x74); /*JE slowpath*/ - addbyte(4+1); - addbyte(0x89); /*MOV -3[EDI+ESI],ECX*/ - addbyte(0x44 | (REG_ECX << 3)); + addbyte(3+1); + addbyte(0x89); /*MOV [EDI+ESI],ECX*/ + addbyte(0x04 | (REG_ECX << 3)); addbyte(REG_EDI | (REG_ESI << 3)); - addbyte(-3); addbyte(0xc3); /*RET*/ addbyte(0x51); /*slowpath: PUSH ECX*/ diff --git a/src/cpu.c b/src/cpu.c index 540ba2d..2a6eda6 100644 --- a/src/cpu.c +++ b/src/cpu.c @@ -68,6 +68,7 @@ int cpu_hasrdtsc; int cpu_hasMMX, cpu_hasMSR; int cpu_hasCR4; int cpu_use_dynarec; +int cpu_cyrix_alignment; uint64_t cpu_CR4_mask; @@ -90,6 +91,7 @@ int timing_iret_rm, timing_iret_v86, timing_iret_pm, timing_iret_pm_outer; int timing_call_rm, timing_call_pm, timing_call_pm_gate, timing_call_pm_gate_inner; int timing_retf_rm, timing_retf_pm, timing_retf_pm_outer; int timing_jmp_rm, timing_jmp_pm, timing_jmp_pm_gate; +int timing_misaligned; static struct { @@ -581,6 +583,9 @@ void cpu_set() } memset(&msr, 0, sizeof(msr)); + + timing_misaligned = 0; + cpu_cyrix_alignment = 0; switch (cpu_s->cpu_type) { @@ -712,6 +717,7 @@ void cpu_set() timing_jmp_rm = 9; timing_jmp_pm = 26; timing_jmp_pm_gate = 37; + timing_misaligned = 3; break; case CPU_486DLC: @@ -744,6 +750,7 @@ void cpu_set() timing_jmp_rm = 9; timing_jmp_pm = 26; timing_jmp_pm_gate = 37; + timing_misaligned = 3; break; case CPU_i486SX: @@ -776,6 +783,7 @@ void cpu_set() timing_jmp_rm = 17; timing_jmp_pm = 19; timing_jmp_pm_gate = 32; + timing_misaligned = 3; break; case CPU_Am486SX: @@ -809,6 +817,7 @@ void cpu_set() timing_jmp_rm = 17; timing_jmp_pm = 19; timing_jmp_pm_gate = 32; + timing_misaligned = 3; break; case CPU_Cx486S: @@ -841,6 +850,7 @@ void cpu_set() timing_jmp_rm = 9; timing_jmp_pm = 26; timing_jmp_pm_gate = 37; + timing_misaligned = 3; break; case CPU_Cx5x86: @@ -872,6 +882,8 @@ void cpu_set() timing_jmp_rm = 5; timing_jmp_pm = 7; timing_jmp_pm_gate = 17; + timing_misaligned = 2; + cpu_cyrix_alignment = 1; break; case CPU_WINCHIP: @@ -909,6 +921,8 @@ void cpu_set() timing_jmp_rm = 5; timing_jmp_pm = 7; timing_jmp_pm_gate = 17; + timing_misaligned = 2; + cpu_cyrix_alignment = 1; codegen_timing_set(&codegen_timing_winchip); break; @@ -942,6 +956,7 @@ void cpu_set() timing_jmp_rm = 3; timing_jmp_pm = 3; timing_jmp_pm_gate = 18; + timing_misaligned = 3; cpu_hasrdtsc = 1; msr.fcr = (1 << 8) | (1 << 9) | (1 << 12) | (1 << 16) | (1 << 19) | (1 << 21); cpu_hasMMX = 0; @@ -981,6 +996,7 @@ void cpu_set() timing_jmp_rm = 3; timing_jmp_pm = 3; timing_jmp_pm_gate = 18; + timing_misaligned = 3; cpu_hasrdtsc = 1; msr.fcr = (1 << 8) | (1 << 9) | (1 << 12) | (1 << 16) | (1 << 19) | (1 << 21); cpu_hasMMX = 1; @@ -1019,6 +1035,8 @@ void cpu_set() timing_jmp_rm = 1; timing_jmp_pm = 4; timing_jmp_pm_gate = 14; + timing_misaligned = 2; + cpu_cyrix_alignment = 1; cpu_hasrdtsc = 1; msr.fcr = (1 << 8) | (1 << 9) | (1 << 12) | (1 << 16) | (1 << 19) | (1 << 21); cpu_hasMMX = 0; @@ -1057,6 +1075,8 @@ void cpu_set() timing_jmp_rm = 1; timing_jmp_pm = 4; timing_jmp_pm_gate = 14; + timing_misaligned = 2; + cpu_cyrix_alignment = 1; cpu_hasrdtsc = 1; msr.fcr = (1 << 8) | (1 << 9) | (1 << 12) | (1 << 16) | (1 << 19) | (1 << 21); cpu_hasMMX = 0; @@ -1129,6 +1149,8 @@ void cpu_set() timing_jmp_rm = 1; timing_jmp_pm = 4; timing_jmp_pm_gate = 14; + timing_misaligned = 2; + cpu_cyrix_alignment = 1; cpu_hasrdtsc = 1; msr.fcr = (1 << 8) | (1 << 9) | (1 << 12) | (1 << 16) | (1 << 19) | (1 << 21); cpu_hasMMX = 1; diff --git a/src/cpu.h b/src/cpu.h index 4a5a679..18671f3 100644 --- a/src/cpu.h +++ b/src/cpu.h @@ -51,6 +51,7 @@ extern int timing_call_rm, timing_call_pm, timing_call_pm_gate, timing_call_pm_g extern int timing_retf_rm, timing_retf_pm, timing_retf_pm_outer; extern int timing_jmp_rm, timing_jmp_pm, timing_jmp_pm_gate; +extern int timing_misaligned; typedef struct { @@ -97,6 +98,8 @@ extern int cpu_iscyrix; extern int cpu_16bitbus; extern int cpu_busspeed; extern int cpu_multi; +/*Cyrix 5x86/6x86 only has data misalignment penalties when crossing 8-byte boundaries*/ +extern int cpu_cyrix_alignment; extern int cpu_hasrdtsc; extern int cpu_hasMSR; diff --git a/src/ibm.h b/src/ibm.h index e2c5b64..20cabea 100644 --- a/src/ibm.h +++ b/src/ibm.h @@ -24,8 +24,8 @@ int writelnext; extern int mmu_perm; #define readmemb(a) ((readlookup2[(a)>>12]==-1)?readmembl(a):*(uint8_t *)(readlookup2[(a) >> 12] + (a))) -#define readmemw(s,a) ((readlookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a))&0xFFF)>0xFFE)?readmemwl(s,a):*(uint16_t *)(readlookup2[(uint32_t)((s)+(a))>>12]+(uint32_t)((s)+(a)))) -#define readmeml(s,a) ((readlookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a))&0xFFF)>0xFFC)?readmemll(s,a):*(uint32_t *)(readlookup2[(uint32_t)((s)+(a))>>12]+(uint32_t)((s)+(a)))) +#define readmemw(s,a) ((readlookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a)) & 1))?readmemwl(s,a):*(uint16_t *)(readlookup2[(uint32_t)((s)+(a))>>12]+(uint32_t)((s)+(a)))) +#define readmeml(s,a) ((readlookup2[(uint32_t)((s)+(a))>>12]==-1 || (s)==0xFFFFFFFF || (((s)+(a)) & 3))?readmemll(s,a):*(uint32_t *)(readlookup2[(uint32_t)((s)+(a))>>12]+(uint32_t)((s)+(a)))) //#define writememb(a,v) if (writelookup2[(a)>>12]==0xFFFFFFFF) writemembl(a,v); else ram[writelookup2[(a)>>12]+((a)&0xFFF)]=v //#define writememw(s,a,v) if (writelookup2[((s)+(a))>>12]==0xFFFFFFFF || (s)==0xFFFFFFFF) writememwl(s,a,v); else *((uint16_t *)(&ram[writelookup2[((s)+(a))>>12]+(((s)+(a))&0xFFF)]))=v diff --git a/src/mem.c b/src/mem.c index ecfc949..b387016 100644 --- a/src/mem.c +++ b/src/mem.c @@ -1099,22 +1099,30 @@ void writememb386l(uint32_t seg, uint32_t addr, uint8_t val) uint16_t readmemwl(uint32_t seg, uint32_t addr) { uint32_t addr2 = mem_logical_addr = seg + addr; - if ((addr2&0xFFF)>0xFFE) - { - if (cr0>>31) - { - if (mmutranslate_read(addr2) == 0xffffffff) return 0xffff; - if (mmutranslate_read(addr2+1) == 0xffffffff) return 0xffff; - } - if (is386) return readmemb386l(seg,addr)|(readmemb386l(seg,addr+1)<<8); - else return readmembl(seg+addr)|(readmembl(seg+addr+1)<<8); - } + if (seg==-1) { x86gpf("NULL segment", 0); printf("NULL segment! rw %04X(%08X):%08X %02X %08X\n",CS,cs,cpu_state.pc,opcode,addr); return -1; } + if (addr2 & 1) + { + if (!cpu_cyrix_alignment || (addr2 & 7) == 7) + cycles -= timing_misaligned; + if ((addr2 & 0xFFF) > 0xFFE) + { + if (cr0 >> 31) + { + if (mmutranslate_read(addr2) == 0xffffffff) return 0xffff; + if (mmutranslate_read(addr2+1) == 0xffffffff) return 0xffff; + } + if (is386) return readmemb386l(seg,addr)|(readmemb386l(seg,addr+1)<<8); + else return readmembl(seg+addr)|(readmembl(seg+addr+1)<<8); + } + else if (readlookup2[addr2 >> 12] != -1) + return *(uint16_t *)(readlookup2[addr2 >> 12] + addr2); + } if (cr0>>31) { addr2 = mmutranslate_read(addr2); @@ -1137,25 +1145,6 @@ uint16_t readmemwl(uint32_t seg, uint32_t addr) void writememwl(uint32_t seg, uint32_t addr, uint16_t val) { uint32_t addr2 = mem_logical_addr = seg + addr; - if ((addr2&0xFFF)>0xFFE) - { - if (cr0>>31) - { - if (mmutranslate_write(addr2) == 0xffffffff) return; - if (mmutranslate_write(addr2+1) == 0xffffffff) return; - } - if (is386) - { - writememb386l(seg,addr,val); - writememb386l(seg,addr+1,val>>8); - } - else - { - writemembl(seg+addr,val); - writemembl(seg+addr+1,val>>8); - } - return; - } if (seg==-1) { @@ -1163,6 +1152,37 @@ void writememwl(uint32_t seg, uint32_t addr, uint16_t val) printf("NULL segment! ww %04X(%08X):%08X %02X %08X\n",CS,cs,cpu_state.pc,opcode,addr); return; } + + if (addr2 & 1) + { + if (!cpu_cyrix_alignment || (addr2 & 7) == 7) + cycles -= timing_misaligned; + if ((addr2 & 0xFFF) > 0xFFE) + { + if (cr0 >> 31) + { + if (mmutranslate_write(addr2) == 0xffffffff) return; + if (mmutranslate_write(addr2+1) == 0xffffffff) return; + } + if (is386) + { + writememb386l(seg,addr,val); + writememb386l(seg,addr+1,val>>8); + } + else + { + writemembl(seg+addr,val); + writemembl(seg+addr+1,val>>8); + } + return; + } + else if (writelookup2[addr2 >> 12] != -1) + { + *(uint16_t *)(writelookup2[addr2 >> 12] + addr2) = val; + return; + } + } + if (page_lookup[addr2>>12]) { page_lookup[addr2>>12]->write_w(addr2, val, page_lookup[addr2>>12]); @@ -1197,15 +1217,6 @@ void writememwl(uint32_t seg, uint32_t addr, uint16_t val) uint32_t readmemll(uint32_t seg, uint32_t addr) { uint32_t addr2 = mem_logical_addr = seg + addr; - if ((addr2&0xFFF)>0xFFC) - { - if (cr0>>31) - { - if (mmutranslate_read(addr2) == 0xffffffff) return 0xffffffff; - if (mmutranslate_read(addr2+3) == 0xffffffff) return 0xffffffff; - } - return readmemwl(seg,addr)|(readmemwl(seg,addr+2)<<16); - } if (seg==-1) { @@ -1213,7 +1224,24 @@ uint32_t readmemll(uint32_t seg, uint32_t addr) printf("NULL segment! rl %04X(%08X):%08X %02X %08X\n",CS,cs,cpu_state.pc,opcode,addr); return -1; } - + + if (addr2 & 3) + { + if (!cpu_cyrix_alignment || (addr2 & 7) > 4) + cycles -= timing_misaligned; + if ((addr2&0xFFF)>0xFFC) + { + if (cr0>>31) + { + if (mmutranslate_read(addr2) == 0xffffffff) return 0xffffffff; + if (mmutranslate_read(addr2+3) == 0xffffffff) return 0xffffffff; + } + return readmemwl(seg,addr)|(readmemwl(seg,addr+2)<<16); + } + else if (readlookup2[addr2 >> 12] != -1) + return *(uint32_t *)(readlookup2[addr2 >> 12] + addr2); + } + if (cr0>>31) { addr2 = mmutranslate_read(addr2); @@ -1236,23 +1264,33 @@ void writememll(uint32_t seg, uint32_t addr, uint32_t val) { uint32_t addr2 = mem_logical_addr = seg + addr; - if ((addr2&0xFFF)>0xFFC) - { - if (cr0>>31) - { - if (mmutranslate_write(addr2) == 0xffffffff) return; - if (mmutranslate_write(addr2+3) == 0xffffffff) return; - } - writememwl(seg,addr,val); - writememwl(seg,addr+2,val>>16); - return; - } if (seg==-1) { x86gpf("NULL segment", 0); printf("NULL segment! wl %04X(%08X):%08X %02X %08X\n",CS,cs,cpu_state.pc,opcode,addr); return; } + if (addr2 & 3) + { + if (!cpu_cyrix_alignment || (addr2 & 7) > 4) + cycles -= timing_misaligned; + if ((addr2 & 0xFFF) > 0xFFC) + { + if (cr0>>31) + { + if (mmutranslate_write(addr2) == 0xffffffff) return; + if (mmutranslate_write(addr2+3) == 0xffffffff) return; + } + writememwl(seg,addr,val); + writememwl(seg,addr+2,val>>16); + return; + } + else if (writelookup2[addr2 >> 12] != -1) + { + *(uint32_t *)(writelookup2[addr2 >> 12] + addr2) = val; + return; + } + } if (page_lookup[addr2>>12]) { page_lookup[addr2>>12]->write_l(addr2, val, page_lookup[addr2>>12]); @@ -1294,15 +1332,6 @@ void writememll(uint32_t seg, uint32_t addr, uint32_t val) uint64_t readmemql(uint32_t seg, uint32_t addr) { uint32_t addr2 = mem_logical_addr = seg + addr; - if ((addr2&0xFFF)>0xFF8) - { - if (cr0>>31) - { - if (mmutranslate_read(addr2) == 0xffffffff) return 0xffffffff; - if (mmutranslate_read(addr2+7) == 0xffffffff) return 0xffffffff; - } - return readmemll(seg,addr)|((uint64_t)readmemll(seg,addr+4)<<32); - } if (seg==-1) { @@ -1310,6 +1339,22 @@ uint64_t readmemql(uint32_t seg, uint32_t addr) printf("NULL segment! rl %04X(%08X):%08X %02X %08X\n",CS,cs,cpu_state.pc,opcode,addr); return -1; } + + if (addr2 & 7) + { + cycles -= timing_misaligned; + if ((addr2 & 0xFFF) > 0xFF8) + { + if (cr0>>31) + { + if (mmutranslate_read(addr2) == 0xffffffff) return 0xffffffff; + if (mmutranslate_read(addr2+7) == 0xffffffff) return 0xffffffff; + } + return readmemll(seg,addr)|((uint64_t)readmemll(seg,addr+4)<<32); + } + else if (readlookup2[addr2 >> 12] != -1) + return *(uint64_t *)(readlookup2[addr2 >> 12] + addr2); + } if (cr0>>31) { @@ -1330,23 +1375,32 @@ void writememql(uint32_t seg, uint32_t addr, uint64_t val) { uint32_t addr2 = mem_logical_addr = seg + addr; - if ((addr2 & 0xFFF) > 0xFF8) - { - if (cr0>>31) - { - if (mmutranslate_write(addr2) == 0xffffffff) return; - if (mmutranslate_write(addr2+7) == 0xffffffff) return; - } - writememll(seg, addr, val); - writememll(seg, addr+4, val >> 32); - return; - } if (seg==-1) { x86gpf("NULL segment", 0); printf("NULL segment! wl %04X(%08X):%08X %02X %08X\n",CS,cs,cpu_state.pc,opcode,addr); return; } + if (addr2 & 7) + { + cycles -= timing_misaligned; + if ((addr2 & 0xFFF) > 0xFF8) + { + if (cr0>>31) + { + if (mmutranslate_write(addr2) == 0xffffffff) return; + if (mmutranslate_write(addr2+7) == 0xffffffff) return; + } + writememll(seg, addr, val); + writememll(seg, addr+4, val >> 32); + return; + } + else if (writelookup2[addr2 >> 12] != -1) + { + *(uint64_t *)(writelookup2[addr2 >> 12] + addr2) = val; + return; + } + } if (page_lookup[addr2>>12]) { page_lookup[addr2>>12]->write_l(addr2, val, page_lookup[addr2>>12]);