Retro/pcem
Watch
1
0
Fork
You've already forked pcem
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Moved FPU register stack into cpu_state structure.

Browse source 
Fixed a couple of FPU bugs in x86-64 recompiler that were introduced in rev 543, fixes lightmap issues in Half-Life.
This commit is contained in:
SarahW 2016-08-25 21:09:02 +01:00
commit 5a02b518b8
17 changed files with 907 additions and 883 deletions
Download patch file Download diff file Expand all files Collapse all files

363
src/codegen_ops_x86-64.h
View file

@ -498,11 +498,11 @@ static void STORE_IMM_REG_L(int reg, uint32_t val)
static void STORE_IMM_ADDR_L(uintptr_t addr, uint32_t val) static void STORE_IMM_ADDR_L(uintptr_t addr, uint32_t val)
{ {
if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x80) if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x100)
{ {
addbyte(0xC7); /*MOVL [addr],val*/ addbyte(0xC7); /*MOVL [addr],val*/
addbyte(0x45); addbyte(0x45);
addbyte(addr - (uintptr_t)&cpu_state); addbyte(addr - (uintptr_t)&cpu_state - 128);
addlong(val); addlong(val);
} }
else if (addr < 0x100000000) else if (addr < 0x100000000)
@ -1572,11 +1572,11 @@ static void STORE_HOST_REG_ADDR_WL(uintptr_t addr, int host_reg)
addbyte(0x0f); /*MOVZX ECX, host_reg*/ addbyte(0x0f); /*MOVZX ECX, host_reg*/
addbyte(0xb7); addbyte(0xb7);
addbyte(0xc0 | (REG_ECX << 3) | (host_reg & 7)); addbyte(0xc0 | (REG_ECX << 3) | (host_reg & 7));
if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x80) if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x100)
{ {
addbyte(0x89); /*MOV addr, ECX*/ addbyte(0x89); /*MOV addr, ECX*/
addbyte(0x45 | (REG_ECX << 3)); addbyte(0x45 | (REG_ECX << 3));
addbyte((uint32_t)addr - (uint32_t)&cpu_state); addbyte((uint32_t)addr - (uint32_t)&cpu_state - 128);
} }
else else
{ {
@ -1588,14 +1588,14 @@ static void STORE_HOST_REG_ADDR_WL(uintptr_t addr, int host_reg)
} }
static void STORE_HOST_REG_ADDR_W(uintptr_t addr, int host_reg) static void STORE_HOST_REG_ADDR_W(uintptr_t addr, int host_reg)
{ {
if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x80) if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x100)
{ {
addbyte(0x66); /*MOVW [addr],host_reg*/ addbyte(0x66); /*MOVW [addr],host_reg*/
if (host_reg & 8) if (host_reg & 8)
addbyte(0x44); addbyte(0x44);
addbyte(0x89); addbyte(0x89);
addbyte(0x45 | ((host_reg & 7) << 3)); addbyte(0x45 | ((host_reg & 7) << 3));
addbyte((uint32_t)addr - (uint32_t)&cpu_state); addbyte((uint32_t)addr - (uint32_t)&cpu_state - 128);
} }
else else
{ {
@ -1610,13 +1610,13 @@ static void STORE_HOST_REG_ADDR_W(uintptr_t addr, int host_reg)
} }
static void STORE_HOST_REG_ADDR(uintptr_t addr, int host_reg) static void STORE_HOST_REG_ADDR(uintptr_t addr, int host_reg)
{ {
if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x80) if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x100)
{ {
if (host_reg & 8) if (host_reg & 8)
addbyte(0x44); addbyte(0x44);
addbyte(0x89); /*MOVL [addr],host_reg*/ addbyte(0x89); /*MOVL [addr],host_reg*/
addbyte(0x45 | ((host_reg & 7) << 3)); addbyte(0x45 | ((host_reg & 7) << 3));
addbyte((uint32_t)addr - (uint32_t)&cpu_state); addbyte((uint32_t)addr - (uint32_t)&cpu_state - 128);
} }
else else
{ {
@ -2681,7 +2681,7 @@ static void LOAD_STACK_TO_EA(int off)
{ {
addbyte(0x8b); /*MOVL EAX,[ESP]*/ addbyte(0x8b); /*MOVL EAX,[ESP]*/
addbyte(0x45 | (REG_EAX << 3)); addbyte(0x45 | (REG_EAX << 3));
addbyte((uint32_t)&ESP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].l));
if (off) if (off)
{ {
addbyte(0x83); /*ADD EAX, off*/ addbyte(0x83); /*ADD EAX, off*/
@ -2694,7 +2694,7 @@ static void LOAD_STACK_TO_EA(int off)
addbyte(0x0f); /*MOVZX EAX,W[ESP]*/ addbyte(0x0f); /*MOVZX EAX,W[ESP]*/
addbyte(0xb7); addbyte(0xb7);
addbyte(0x45 | (REG_EAX << 3)); addbyte(0x45 | (REG_EAX << 3));
addbyte((uint32_t)&ESP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].w));
if (off) if (off)
{ {
addbyte(0x66); /*ADD AX, off*/ addbyte(0x66); /*ADD AX, off*/
@ -2709,7 +2709,7 @@ static void LOAD_EBP_TO_EA(int off)
{ {
addbyte(0x8b); /*MOVL EAX,[EBP]*/ addbyte(0x8b); /*MOVL EAX,[EBP]*/
addbyte(0x45 | (REG_EAX << 3)); addbyte(0x45 | (REG_EAX << 3));
addbyte((uint32_t)&EBP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_EBP].l));
if (off) if (off)
{ {
addbyte(0x83); /*ADD EAX, off*/ addbyte(0x83); /*ADD EAX, off*/
@ -2722,7 +2722,7 @@ static void LOAD_EBP_TO_EA(int off)
addbyte(0x0f); /*MOVZX EAX,W[EBP]*/ addbyte(0x0f); /*MOVZX EAX,W[EBP]*/
addbyte(0xb7); addbyte(0xb7);
addbyte(0x45 | (REG_EAX << 3)); addbyte(0x45 | (REG_EAX << 3));
addbyte((uint32_t)&EBP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_BP].l));
if (off) if (off)
{ {
addbyte(0x66); /*ADD AX, off*/ addbyte(0x66); /*ADD AX, off*/
@ -2740,14 +2740,14 @@ static void SP_MODIFY(int off)
{ {
addbyte(0x83); /*ADD [ESP], off*/ addbyte(0x83); /*ADD [ESP], off*/
addbyte(0x45); addbyte(0x45);
addbyte((uint32_t)&ESP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].l));
addbyte(off); addbyte(off);
} }
else else
{ {
addbyte(0x81); /*ADD [ESP], off*/ addbyte(0x81); /*ADD [ESP], off*/
addbyte(0x45); addbyte(0x45);
addbyte((uint32_t)&ESP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].l));
addlong(off); addlong(off);
} }
} }
@ -2758,7 +2758,7 @@ static void SP_MODIFY(int off)
addbyte(0x66); /*ADD [SP], off*/ addbyte(0x66); /*ADD [SP], off*/
addbyte(0x83); addbyte(0x83);
addbyte(0x45); addbyte(0x45);
addbyte((uint32_t)&SP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].w));
addbyte(off); addbyte(off);
} }
else else
@ -2766,7 +2766,7 @@ static void SP_MODIFY(int off)
addbyte(0x66); /*ADD [SP], off*/ addbyte(0x66); /*ADD [SP], off*/
addbyte(0x81); addbyte(0x81);
addbyte(0x45); addbyte(0x45);
addbyte((uint32_t)&SP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].w));
addword(off); addword(off);
} }
} }
@ -3348,70 +3348,73 @@ static void FP_FXCH(int reg)
addbyte(0x8b); /*MOV EAX, [TOP]*/ addbyte(0x8b); /*MOV EAX, [TOP]*/
addbyte(0x45); addbyte(0x45);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0x48); /*MOV RSI, ST*/
addbyte(0xbe);
addquad((uint64_t)ST);
addbyte(0x89); /*MOV EBX, EAX*/ addbyte(0x89); /*MOV EBX, EAX*/
addbyte(0xc3); addbyte(0xc3);
addbyte(0x83); /*ADD EAX, reg*/ addbyte(0x83); /*ADD EAX, reg*/
addbyte(0xc0); addbyte(0xc0);
addbyte(reg); addbyte(reg);
addbyte(0x48); /*MOV RDX, [RSI + RBX*8]*/ addbyte(0x48); /*MOV RDX, ST[RBX*8]*/
addbyte(0x8b); addbyte(0x8b);
addbyte(0x14); addbyte(0x54);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
addbyte(0x83); /*AND EAX, 7*/ addbyte(0x83); /*AND EAX, 7*/
addbyte(0xe0); addbyte(0xe0);
addbyte(0x07); addbyte(0x07);
addbyte(0x48); /*MOV RCX, [RSI + RAX*8]*/ addbyte(0x48); /*MOV RCX, ST[RAX*8]*/
addbyte(0x8b); addbyte(0x8b);
addbyte(0x0c); addbyte(0x4c);
addbyte(0xc6); addbyte(0xc5);
addbyte(0x48); /*MOV [RSI + RAX*8], RDX*/ addbyte(cpu_state_offset(ST));
addbyte(0x48); /*MOV ST[RAX*8], RDX*/
addbyte(0x89); addbyte(0x89);
addbyte(0x14); addbyte(0x54);
addbyte(0xc6); addbyte(0xc5);
addbyte(0x48); /*MOV [RSI + RBX*8], RCX*/ addbyte(cpu_state_offset(ST));
addbyte(0x48); /*MOV ST[RBX*8], RCX*/
addbyte(0x89); addbyte(0x89);
addbyte(0x0c); addbyte(0x4c);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
addbyte(0x48); /*MOVL ESI, tag*/
addbyte(0xbe);
addquad((uintptr_t)cpu_state.tag);
addbyte(0x8a); /*MOV CL, tag[EAX]*/ addbyte(0x8a); /*MOV CL, tag[EAX]*/
addbyte(0x14); addbyte(0x4c);
addbyte(0x1e); addbyte(0x05);
addbyte(cpu_state_offset(tag));
addbyte(0x8a); /*MOV DL, tag[EBX]*/ addbyte(0x8a); /*MOV DL, tag[EBX]*/
addbyte(0x0c); addbyte(0x54);
addbyte(0x06); addbyte(0x1d);
addbyte(cpu_state_offset(tag));
addbyte(0x88); /*MOV tag[EBX], CL*/ addbyte(0x88); /*MOV tag[EBX], CL*/
addbyte(0x14); addbyte(0x4c);
addbyte(0x06); addbyte(0x1d);
addbyte(cpu_state_offset(tag));
addbyte(0x88); /*MOV tag[EAX], DL*/ addbyte(0x88); /*MOV tag[EAX], DL*/
addbyte(0x0c); addbyte(0x54);
addbyte(0x1e); addbyte(0x05);
addbyte(cpu_state_offset(tag));
addbyte(0x48); /*MOV RSI, ST_i64*/ addbyte(0x48); /*MOV RDX, MM[RBX*8]*/
addbyte(0xbe);
addquad((uint64_t)MM);
addbyte(0x48); /*MOV RDX, [RSI + RBX*8]*/
addbyte(0x8b); addbyte(0x8b);
addbyte(0x14); addbyte(0x54);
addbyte(0xde); addbyte(0xdd);
addbyte(0x48); /*MOV RCX, [RSI + RAX*8]*/ addbyte(cpu_state_offset(MM));
addbyte(0x48); /*MOV RCX, MM[RAX*8]*/
addbyte(0x8b); addbyte(0x8b);
addbyte(0x0c); addbyte(0x4c);
addbyte(0xc6); addbyte(0xc5);
addbyte(0x48); /*MOV [RSI + RAX*8], RDX*/ addbyte(cpu_state_offset(MM));
addbyte(0x48); /*MOV MM[RAX*8], RDX*/
addbyte(0x89); addbyte(0x89);
addbyte(0x14); addbyte(0x54);
addbyte(0xc6); addbyte(0xc5);
addbyte(0x48); /*MOV [RSI + RBX*8], RCX*/ addbyte(cpu_state_offset(MM));
addbyte(0x48); /*MOV MM[RBX*8], RCX*/
addbyte(0x89); addbyte(0x89);
addbyte(0x0c); addbyte(0x4c);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(MM));
reg = reg; reg = reg;
} }
@ -3445,31 +3448,31 @@ static void FP_FLD(int reg)
addbyte(0x48); /*MOV RCX, ST[EAX*8]*/ addbyte(0x48); /*MOV RCX, ST[EAX*8]*/
addbyte(0x8b); addbyte(0x8b);
addbyte(0x0c); addbyte(0x4c);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x83); /*AND EBX, 7*/ addbyte(0x83); /*AND EBX, 7*/
addbyte(0xe3); addbyte(0xe3);
addbyte(0x07); addbyte(0x07);
addbyte(0x48); /*MOV RDX, ST_i64[EAX*8]*/ addbyte(0x48); /*MOV RDX, ST_i64[EAX*8]*/
addbyte(0x8b); addbyte(0x8b);
addbyte(0x14); addbyte(0x54);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)MM); addbyte(cpu_state_offset(MM));
addbyte(0x8a); /*MOV AL, [tag+EAX]*/ addbyte(0x8a); /*MOV AL, [tag+EAX]*/
addbyte(0x44); addbyte(0x44);
addbyte(0x1d); addbyte(0x05);
addbyte(cpu_state_offset(tag)); addbyte(cpu_state_offset(tag));
addbyte(0x48); /*MOV ST[EBX*8], RCX*/ addbyte(0x48); /*MOV ST[EBX*8], RCX*/
addbyte(0x89); addbyte(0x89);
addbyte(0x0c); addbyte(0x4c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x48); /*MOV ST_i64[EBX*8], RDX*/ addbyte(0x48); /*MOV ST_i64[EBX*8], RDX*/
addbyte(0x89); addbyte(0x89);
addbyte(0x14); addbyte(0x54);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)MM); addbyte(cpu_state_offset(MM));
addbyte(0x88); /*MOV [tag+EBX], AL*/ addbyte(0x88); /*MOV [tag+EBX], AL*/
addbyte(0x44); addbyte(0x44);
addbyte(0x1d); addbyte(0x1d);
@ -3487,9 +3490,9 @@ static void FP_FST(int reg)
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0x48); /*MOV RCX, ST[EAX*8]*/ addbyte(0x48); /*MOV RCX, ST[EAX*8]*/
addbyte(0x8b); addbyte(0x8b);
addbyte(0x0c); addbyte(0x4c);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x8a); /*MOV BL, [tag+EAX]*/ addbyte(0x8a); /*MOV BL, [tag+EAX]*/
addbyte(0x5c); addbyte(0x5c);
addbyte(0x05); addbyte(0x05);
@ -3507,9 +3510,9 @@ static void FP_FST(int reg)
addbyte(0x48); /*MOV ST[EAX*8], RCX*/ addbyte(0x48); /*MOV ST[EAX*8], RCX*/
addbyte(0x89); addbyte(0x89);
addbyte(0x0c); addbyte(0x4c);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x88); /*MOV [tag+EAX], BL*/ addbyte(0x88); /*MOV [tag+EAX], BL*/
addbyte(0x5c); addbyte(0x5c);
addbyte(0x05); addbyte(0x05);
@ -3564,9 +3567,9 @@ static void FP_LOAD_S()
addbyte(0x66); /*MOVQ [ST+EBX*8], XMM0*/ addbyte(0x66); /*MOVQ [ST+EBX*8], XMM0*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0xd6); addbyte(0xd6);
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x0f); /*SETE [tag+EBX]*/ addbyte(0x0f); /*SETE [tag+EBX]*/
addbyte(0x94); addbyte(0x94);
addbyte(0x44); addbyte(0x44);
@ -3592,9 +3595,9 @@ static void FP_LOAD_D()
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0x48); /*MOVQ [ST+EBX*8], RAX*/ addbyte(0x48); /*MOVQ [ST+EBX*8], RAX*/
addbyte(0x89); addbyte(0x89);
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x0f); /*SETE [tag+EBX]*/ addbyte(0x0f); /*SETE [tag+EBX]*/
addbyte(0x94); addbyte(0x94);
addbyte(0x44); addbyte(0x44);
@ -3628,9 +3631,9 @@ static void FP_LOAD_IW()
addbyte(0x66); /*MOVQ [ST+EBX*8], XMM0*/ addbyte(0x66); /*MOVQ [ST+EBX*8], XMM0*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0xd6); addbyte(0xd6);
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x0f); /*SETE [tag+EBX]*/ addbyte(0x0f); /*SETE [tag+EBX]*/
addbyte(0x94); addbyte(0x94);
addbyte(0x44); addbyte(0x44);
@ -3660,9 +3663,9 @@ static void FP_LOAD_IL()
addbyte(0x66); /*MOVQ [ST+EBX*8], XMM0*/ addbyte(0x66); /*MOVQ [ST+EBX*8], XMM0*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0xd6); addbyte(0xd6);
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x0f); /*SETE [tag+EBX]*/ addbyte(0x0f); /*SETE [tag+EBX]*/
addbyte(0x94); addbyte(0x94);
addbyte(0x44); addbyte(0x44);
@ -3690,9 +3693,9 @@ static void FP_LOAD_IQ()
addbyte(0xc0); addbyte(0xc0);
addbyte(0x48); /*MOV [ST_i64+EBX*8], RAX*/ addbyte(0x48); /*MOV [ST_i64+EBX*8], RAX*/
addbyte(0x89); addbyte(0x89);
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)MM); addbyte(cpu_state_offset(MM));
addbyte(0x89); /*MOV TOP, EBX*/ addbyte(0x89); /*MOV TOP, EBX*/
addbyte(0x5d); addbyte(0x5d);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
@ -3702,9 +3705,9 @@ static void FP_LOAD_IQ()
addbyte(0x66); /*MOVQ [ST+EBX*8], XMM0*/ addbyte(0x66); /*MOVQ [ST+EBX*8], XMM0*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0xd6); addbyte(0xd6);
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x0c); /*OR AL, TAG_UINT64*/ addbyte(0x0c); /*OR AL, TAG_UINT64*/
addbyte(TAG_UINT64); addbyte(TAG_UINT64);
addbyte(0x88); /*MOV [tag+EBX], AL*/ addbyte(0x88); /*MOV [tag+EBX], AL*/
@ -3730,9 +3733,9 @@ static int FP_LOAD_REG(int reg)
addbyte(0xf3); /*MOVQ XMM0, ST[EBX*8]*/ addbyte(0xf3); /*MOVQ XMM0, ST[EBX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0xf2); /*CVTSD2SS XMM0, XMM0*/ addbyte(0xf2); /*CVTSD2SS XMM0, XMM0*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x5a); addbyte(0x5a);
@ -3760,9 +3763,9 @@ static void FP_LOAD_REG_D(int reg, int *host_reg1, int *host_reg2)
} }
addbyte(0x48); /*MOV RBX, ST[EBX*8]*/ addbyte(0x48); /*MOV RBX, ST[EBX*8]*/
addbyte(0x8b); addbyte(0x8b);
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
*host_reg1 = REG_EBX; *host_reg1 = REG_EBX;
} }
@ -3809,9 +3812,9 @@ static int FP_LOAD_REG_INT_W(int reg)
addbyte(0xf3); /*MOVQ XMM0, ST[EAX*8]*/ addbyte(0xf3); /*MOVQ XMM0, ST[EAX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x04); addbyte(0x44);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
CALL_FUNC(x87_fround); CALL_FUNC(x87_fround);
@ -3839,9 +3842,9 @@ static int FP_LOAD_REG_INT(int reg)
addbyte(0xf3); /*MOVQ XMM0, ST[EAX*8]*/ addbyte(0xf3); /*MOVQ XMM0, ST[EAX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x04); addbyte(0x44);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
CALL_FUNC(x87_fround); CALL_FUNC(x87_fround);
@ -3874,9 +3877,9 @@ static void FP_LOAD_REG_INT_Q(int reg, int *host_reg1, int *host_reg2)
and just load the 64-bit integer representation directly */ and just load the 64-bit integer representation directly */
addbyte(0x48); /*MOV RAX, [ST_i64+EAX*8]*/ addbyte(0x48); /*MOV RAX, [ST_i64+EAX*8]*/
addbyte(0x8b); addbyte(0x8b);
addbyte(0x04); addbyte(0x44);
addbyte(0xc5); addbyte(0xc5);
addlong((uint32_t)MM); addbyte(cpu_state_offset(MM));
addbyte(0x48); /*XCHG RBX, RAX*/ addbyte(0x48); /*XCHG RBX, RAX*/
addbyte(0x93); addbyte(0x93);
@ -3893,23 +3896,23 @@ static void FP_LOAD_REG_INT_Q(int reg, int *host_reg1, int *host_reg2)
addbyte(TAG_UINT64); addbyte(TAG_UINT64);
addbyte(0x74); /*JZ +*/ addbyte(0x74); /*JZ +*/
addbyte(8+2); addbyte(5+2);
addbyte(0x48); /*MOV RAX, [ST_i64+EAX*8]*/ addbyte(0x48); /*MOV RAX, [ST_i64+EAX*8]*/
addbyte(0x8b); addbyte(0x8b);
addbyte(0x04); addbyte(0x44);
addbyte(0xc5); addbyte(0xc5);
addlong((uint32_t)MM); addbyte(cpu_state_offset(MM));
addbyte(0xeb); /*JMP done*/ addbyte(0xeb); /*JMP done*/
addbyte(9+12); addbyte(6+12);
addbyte(0xf3); /*MOVQ XMM0, ST[EAX*8]*/ addbyte(0xf3); /*MOVQ XMM0, ST[EAX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x04); addbyte(0x44);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
CALL_FUNC(x87_fround); CALL_FUNC(x87_fround);
@ -3931,9 +3934,6 @@ static void FP_OP_REG(int op, int dst, int src)
addbyte(0x8b); /*MOV EAX, [TOP]*/ addbyte(0x8b); /*MOV EAX, [TOP]*/
addbyte(0x45); addbyte(0x45);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0x48); /*MOV RSI, ST*/
addbyte(0xbe);
addquad((uint64_t)ST);
addbyte(0x89); /*MOV EBX, EAX*/ addbyte(0x89); /*MOV EBX, EAX*/
addbyte(0xc3); addbyte(0xc3);
if (dst) if (dst)
@ -3954,72 +3954,86 @@ static void FP_OP_REG(int op, int dst, int src)
addbyte(0xe0 | REG_EBX); addbyte(0xe0 | REG_EBX);
addbyte(0x07); addbyte(0x07);
} }
addbyte(0x80); /*AND tag[EAX], ~TAG_UINT64*/
addbyte(0x64);
addbyte(0x05);
addbyte(cpu_state_offset(tag));
addbyte(~TAG_UINT64);
if (op == FPU_DIVR || op == FPU_SUBR) if (op == FPU_DIVR || op == FPU_SUBR)
{ {
addbyte(0xf3); /*MOVQ XMM0, [RSI+RBX*8]*/ addbyte(0xf3); /*MOVQ XMM0, ST[RBX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x04); addbyte(0x44);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
} }
else else
{ {
addbyte(0xf3); /*MOVQ XMM0, [RSI+RAX*8]*/ addbyte(0xf3); /*MOVQ XMM0, ST[RAX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x04); addbyte(0x44);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
} }
switch (op) switch (op)
{ {
case FPU_ADD: case FPU_ADD:
addbyte(0xf2); /*ADDSD XMM0, [RSI+RBX*8]*/ addbyte(0xf2); /*ADDSD XMM0, ST[RBX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x58); addbyte(0x58);
addbyte(0x04); addbyte(0x44);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
break; break;
case FPU_DIV: case FPU_DIV:
addbyte(0xf2); /*DIVSD XMM0, [RSI+RBX*8]*/ addbyte(0xf2); /*DIVSD XMM0, ST[RBX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x5e); addbyte(0x5e);
addbyte(0x04); addbyte(0x44);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
break; break;
case FPU_DIVR: case FPU_DIVR:
addbyte(0xf2); /*DIVSD XMM0, [RSI+RAX*8]*/ addbyte(0xf2); /*DIVSD XMM0, ST[RAX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x5e); addbyte(0x5e);
addbyte(0x04); addbyte(0x44);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
break; break;
case FPU_MUL: case FPU_MUL:
addbyte(0xf2); /*MULSD XMM0, [RSI+RBX*8]*/ addbyte(0xf2); /*MULSD XMM0, ST[RBX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x59); addbyte(0x59);
addbyte(0x04); addbyte(0x44);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
break; break;
case FPU_SUB: case FPU_SUB:
addbyte(0xf2); /*SUBSD XMM0, [RSI+RBX*8]*/ addbyte(0xf2); /*SUBSD XMM0, ST[RBX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x5c); addbyte(0x5c);
addbyte(0x04); addbyte(0x44);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
break; break;
case FPU_SUBR: case FPU_SUBR:
addbyte(0xf2); /*SUBSD XMM0, [RSI+RAX*8]*/ addbyte(0xf2); /*SUBSD XMM0, ST[RAX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x5c); addbyte(0x5c);
addbyte(0x04); addbyte(0x44);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
break; break;
} }
addbyte(0x66); /*MOVQ [RSI+RAX*8], XMM0*/ addbyte(0x66); /*MOVQ [RSI+RAX*8], XMM0*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0xd6); addbyte(0xd6);
addbyte(0x04); addbyte(0x44);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
} }
static void FP_OP_MEM(int op) static void FP_OP_MEM(int op)
@ -4027,14 +4041,12 @@ static void FP_OP_MEM(int op)
addbyte(0x8b); /*MOV EAX, [TOP]*/ addbyte(0x8b); /*MOV EAX, [TOP]*/
addbyte(0x45); addbyte(0x45);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0x48); /*MOV RSI, ST*/ addbyte(0xf3); /*MOVQ XMM0, ST[RAX*8]*/
addbyte(0xbe);
addquad((uint64_t)ST);
addbyte(0xf3); /*MOVQ XMM0, [RSI+RAX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x04); addbyte(0x44);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
addbyte(0x80); /*AND tag[EAX], ~TAG_UINT64*/ addbyte(0x80); /*AND tag[EAX], ~TAG_UINT64*/
addbyte(0x64); addbyte(0x64);
addbyte(0x05); addbyte(0x05);
@ -4082,19 +4094,21 @@ static void FP_OP_MEM(int op)
} }
if (op == FPU_DIVR || op == FPU_SUBR) if (op == FPU_DIVR || op == FPU_SUBR)
{ {
addbyte(0x66); /*MOVQ [RSI+RAX*8], XMM1*/ addbyte(0x66); /*MOVQ ST[RAX*8], XMM1*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0xd6); addbyte(0xd6);
addbyte(0x0c); addbyte(0x4c);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
} }
else else
{ {
addbyte(0x66); /*MOVQ [RSI+RAX*8], XMM0*/ addbyte(0x66); /*MOVQ ST[RAX*8], XMM0*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0xd6); addbyte(0xd6);
addbyte(0x04); addbyte(0x44);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
} }
} }
@ -4182,9 +4196,6 @@ static void FP_COMPARE_REG(int dst, int src)
addbyte(0x8b); /*MOV EAX, [TOP]*/ addbyte(0x8b); /*MOV EAX, [TOP]*/
addbyte(0x45); addbyte(0x45);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0x48); /*MOV RSI, ST*/
addbyte(0xbe);
addquad((uint64_t)ST);
addbyte(0x89); /*MOV EBX, EAX*/ addbyte(0x89); /*MOV EBX, EAX*/
addbyte(0xc3); addbyte(0xc3);
if (src || dst) if (src || dst)
@ -4209,29 +4220,33 @@ static void FP_COMPARE_REG(int dst, int src)
if (src) if (src)
{ {
addbyte(0xf3); /*MOVQ XMM0, [RSI+RBX*8]*/ addbyte(0xf3); /*MOVQ XMM0, ST[RBX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x04); addbyte(0x44);
addbyte(0xde); addbyte(0xdd);
addbyte(0x66); /*COMISD XMM0, [RSI+RAX*8]*/ addbyte(cpu_state_offset(ST));
addbyte(0x66); /*COMISD XMM0, ST[RAX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x2f); addbyte(0x2f);
addbyte(0x04); addbyte(0x44);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
} }
else else
{ {
addbyte(0xf3); /*MOVQ XMM0, [RSI+RAX*8]*/ addbyte(0xf3); /*MOVQ XMM0, ST[RAX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x04); addbyte(0x44);
addbyte(0xc6); addbyte(0xc5);
addbyte(0x66); /*COMISD XMM0, [RSI+RBX*8]*/ addbyte(cpu_state_offset(ST));
addbyte(0x66); /*COMISD XMM0, ST[RBX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x2f); addbyte(0x2f);
addbyte(0x04); addbyte(0x44);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
} }
addbyte(0x9f); /*LAHF*/ addbyte(0x9f); /*LAHF*/
@ -4251,9 +4266,6 @@ static void FP_COMPARE_MEM()
addbyte(0x8b); /*MOV EAX, [TOP]*/ addbyte(0x8b); /*MOV EAX, [TOP]*/
addbyte(0x45); addbyte(0x45);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0x48); /*MOV RSI, ST*/
addbyte(0xbe);
addquad((uint64_t)ST);
addbyte(0x8a); /*MOV CL, [npxs+1]*/ addbyte(0x8a); /*MOV CL, [npxs+1]*/
addbyte(0x0c); addbyte(0x0c);
@ -4261,11 +4273,12 @@ static void FP_COMPARE_MEM()
addlong(((uintptr_t)&npxs) + 1); addlong(((uintptr_t)&npxs) + 1);
// addbyte(0xdb); /*FCLEX*/ // addbyte(0xdb); /*FCLEX*/
// addbyte(0xe2); // addbyte(0xe2);
addbyte(0xf3); /*MOVQ XMM0, [RSI+RAX*8]*/ addbyte(0xf3); /*MOVQ XMM0, ST[RAX*8]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x04); addbyte(0x44);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
addbyte(0x80); /*AND CL, ~(C0|C2|C3)*/ addbyte(0x80); /*AND CL, ~(C0|C2|C3)*/
addbyte(0xe1); addbyte(0xe1);
addbyte((~(C0|C2|C3)) >> 8); addbyte((~(C0|C2|C3)) >> 8);
@ -4415,9 +4428,9 @@ static int LOAD_MMX_D(int guest_reg)
int host_reg = REG_EBX; int host_reg = REG_EBX;
addbyte(0x8b); /*MOV EBX, reg*/ addbyte(0x8b); /*MOV EBX, reg*/
addbyte(0x04 | (host_reg << 3)); addbyte(0x44 | (host_reg << 3));
addbyte(0x25); addbyte(0x25);
addlong((uint32_t)&MM[guest_reg].l[0]); addbyte(cpu_state_offset(MM[guest_reg].l[0]));
return host_reg; return host_reg;
} }
@ -4430,9 +4443,9 @@ static void LOAD_MMX_Q(int guest_reg, int *host_reg1, int *host_reg2)
else else
addbyte(0x48); addbyte(0x48);
addbyte(0x8b); /*MOV RBX, reg*/ addbyte(0x8b); /*MOV RBX, reg*/
addbyte(0x04 | ((host_reg & 7) << 3)); addbyte(0x44 | ((host_reg & 7) << 3));
addbyte(0x25); addbyte(0x25);
addlong((uint32_t)&MM[guest_reg].q); addbyte(cpu_state_offset(MM[guest_reg].q));
*host_reg1 = host_reg; *host_reg1 = host_reg;
} }
@ -4444,9 +4457,9 @@ static int LOAD_MMX_Q_MMX(int guest_reg)
addbyte(0xf3); /*MOV XMMx, reg*/ addbyte(0xf3); /*MOV XMMx, reg*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x04 | ((dst_reg & 7) << 3)); addbyte(0x44 | ((dst_reg & 7) << 3));
addbyte(0x25); addbyte(0x25);
addlong((uint32_t)&MM[guest_reg].q); addbyte(cpu_state_offset(MM[guest_reg].q));
return dst_reg; return dst_reg;
} }
@ -4471,16 +4484,16 @@ static int LOAD_INT_TO_MMX(int src_reg1, int src_reg2)
static void STORE_MMX_LQ(int guest_reg, int host_reg1) static void STORE_MMX_LQ(int guest_reg, int host_reg1)
{ {
addbyte(0xC7); /*MOVL [reg],0*/ addbyte(0xC7); /*MOVL [reg],0*/
addbyte(0x04); addbyte(0x44);
addbyte(0x25); addbyte(0x25);
addlong((uint32_t)&MM[guest_reg].l[1]); addbyte(cpu_state_offset(MM[guest_reg].l[1]));
addlong(0); addlong(0);
if (host_reg1 & 8) if (host_reg1 & 8)
addbyte(0x44); addbyte(0x44);
addbyte(0x89); /*MOVL [reg],host_reg*/ addbyte(0x89); /*MOVL [reg],host_reg*/
addbyte(0x04 | ((host_reg1 & 7) << 3)); addbyte(0x44 | ((host_reg1 & 7) << 3));
addbyte(0x25); addbyte(0x25);
addlong((uint32_t)&MM[guest_reg].l[0]); addbyte(cpu_state_offset(MM[guest_reg].l[0]));
} }
static void STORE_MMX_Q(int guest_reg, int host_reg1, int host_reg2) static void STORE_MMX_Q(int guest_reg, int host_reg1, int host_reg2)
{ {
@ -4489,18 +4502,18 @@ static void STORE_MMX_Q(int guest_reg, int host_reg1, int host_reg2)
else else
addbyte(0x48); addbyte(0x48);
addbyte(0x89); /*MOV [reg],host_reg*/ addbyte(0x89); /*MOV [reg],host_reg*/
addbyte(0x04 | ((host_reg1 & 7) << 3)); addbyte(0x44 | ((host_reg1 & 7) << 3));
addbyte(0x25); addbyte(0x25);
addlong((uint32_t)&MM[guest_reg].l[0]); addbyte(cpu_state_offset(MM[guest_reg].l[0]));
} }
static void STORE_MMX_Q_MMX(int guest_reg, int host_reg) static void STORE_MMX_Q_MMX(int guest_reg, int host_reg)
{ {
addbyte(0x66); /*MOVQ [guest_reg],host_reg*/ addbyte(0x66); /*MOVQ [guest_reg],host_reg*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0xd6); addbyte(0xd6);
addbyte(0x04 | (host_reg << 3)); addbyte(0x44 | (host_reg << 3));
addbyte(0x25); addbyte(0x25);
addlong((uint32_t)&MM[guest_reg].q); addbyte(cpu_state_offset(MM[guest_reg].q));
} }
#define MMX_x86_OP(name, opcode) \ #define MMX_x86_OP(name, opcode) \

304
src/codegen_ops_x86.h
View file

@ -52,11 +52,11 @@ static void STORE_IMM_ADDR_W(uintptr_t addr, uint16_t val)
} }
static void STORE_IMM_ADDR_L(uintptr_t addr, uint32_t val) static void STORE_IMM_ADDR_L(uintptr_t addr, uint32_t val)
{ {
if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x80) if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x100)
{ {
addbyte(0xC7); /*MOVL [addr],val*/ addbyte(0xC7); /*MOVL [addr],val*/
addbyte(0x45); addbyte(0x45);
addbyte(addr - (uint32_t)&cpu_state); addbyte(addr - (uint32_t)&cpu_state - 128);
addlong(val); addlong(val);
} }
else else
@ -239,12 +239,12 @@ static void RELEASE_REG(int host_reg)
static void STORE_HOST_REG_ADDR_W(uintptr_t addr, int host_reg) static void STORE_HOST_REG_ADDR_W(uintptr_t addr, int host_reg)
{ {
if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x80) if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x100)
{ {
addbyte(0x66); /*MOVW [addr],host_reg*/ addbyte(0x66); /*MOVW [addr],host_reg*/
addbyte(0x89); addbyte(0x89);
addbyte(0x45 | (host_reg << 3)); addbyte(0x45 | (host_reg << 3));
addbyte((uint32_t)addr - (uint32_t)&cpu_state); addbyte((uint32_t)addr - (uint32_t)&cpu_state - 128);
} }
else else
{ {
@ -256,11 +256,11 @@ static void STORE_HOST_REG_ADDR_W(uintptr_t addr, int host_reg)
} }
static void STORE_HOST_REG_ADDR(uintptr_t addr, int host_reg) static void STORE_HOST_REG_ADDR(uintptr_t addr, int host_reg)
{ {
if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x80) if (addr >= (uintptr_t)&cpu_state && addr < ((uintptr_t)&cpu_state)+0x100)
{ {
addbyte(0x89); /*MOVL [addr],host_reg*/ addbyte(0x89); /*MOVL [addr],host_reg*/
addbyte(0x45 | (host_reg << 3)); addbyte(0x45 | (host_reg << 3));
addbyte((uint32_t)addr - (uint32_t)&cpu_state); addbyte((uint32_t)addr - (uint32_t)&cpu_state - 128);
} }
else else
{ {
@ -1359,7 +1359,7 @@ static void LOAD_STACK_TO_EA(int off)
{ {
addbyte(0x8b); /*MOVL EAX,[ESP]*/ addbyte(0x8b); /*MOVL EAX,[ESP]*/
addbyte(0x45 | (REG_EAX << 3)); addbyte(0x45 | (REG_EAX << 3));
addbyte((uint32_t)&ESP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].l));
if (off) if (off)
{ {
addbyte(0x83); /*ADD EAX, off*/ addbyte(0x83); /*ADD EAX, off*/
@ -1372,7 +1372,7 @@ static void LOAD_STACK_TO_EA(int off)
addbyte(0x0f); /*MOVZX EAX,W[ESP]*/ addbyte(0x0f); /*MOVZX EAX,W[ESP]*/
addbyte(0xb7); addbyte(0xb7);
addbyte(0x45 | (REG_EAX << 3)); addbyte(0x45 | (REG_EAX << 3));
addbyte((uint32_t)&ESP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].w));
if (off) if (off)
{ {
addbyte(0x66); /*ADD AX, off*/ addbyte(0x66); /*ADD AX, off*/
@ -1388,7 +1388,7 @@ static void LOAD_EBP_TO_EA(int off)
{ {
addbyte(0x8b); /*MOVL EAX,[EBP]*/ addbyte(0x8b); /*MOVL EAX,[EBP]*/
addbyte(0x45 | (REG_EAX << 3)); addbyte(0x45 | (REG_EAX << 3));
addbyte((uint32_t)&EBP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_EBP].l));
if (off) if (off)
{ {
addbyte(0x83); /*ADD EAX, off*/ addbyte(0x83); /*ADD EAX, off*/
@ -1401,7 +1401,7 @@ static void LOAD_EBP_TO_EA(int off)
addbyte(0x0f); /*MOVZX EAX,W[EBP]*/ addbyte(0x0f); /*MOVZX EAX,W[EBP]*/
addbyte(0xb7); addbyte(0xb7);
addbyte(0x45 | (REG_EAX << 3)); addbyte(0x45 | (REG_EAX << 3));
addbyte((uint32_t)&EBP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_EBP].w));
if (off) if (off)
{ {
addbyte(0x66); /*ADD AX, off*/ addbyte(0x66); /*ADD AX, off*/
@ -1419,14 +1419,14 @@ static void SP_MODIFY(int off)
{ {
addbyte(0x83); /*ADD [ESP], off*/ addbyte(0x83); /*ADD [ESP], off*/
addbyte(0x45); addbyte(0x45);
addbyte((uint32_t)&ESP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].l));
addbyte(off); addbyte(off);
} }
else else
{ {
addbyte(0x81); /*ADD [ESP], off*/ addbyte(0x81); /*ADD [ESP], off*/
addbyte(0x45); addbyte(0x45);
addbyte((uint32_t)&ESP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].l));
addlong(off); addlong(off);
} }
} }
@ -1437,7 +1437,7 @@ static void SP_MODIFY(int off)
addbyte(0x66); /*ADD [SP], off*/ addbyte(0x66); /*ADD [SP], off*/
addbyte(0x83); addbyte(0x83);
addbyte(0x45); addbyte(0x45);
addbyte((uint32_t)&SP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].w));
addbyte(off); addbyte(off);
} }
else else
@ -1445,7 +1445,7 @@ static void SP_MODIFY(int off)
addbyte(0x66); /*ADD [SP], off*/ addbyte(0x66); /*ADD [SP], off*/
addbyte(0x81); addbyte(0x81);
addbyte(0x45); addbyte(0x45);
addbyte((uint32_t)&SP - (uint32_t)&EAX); addbyte(cpu_state_offset(regs[REG_ESP].w));
addword(off); addword(off);
} }
} }
@ -1600,7 +1600,7 @@ static int BRANCH_COND_BE(int pc_offset, uint32_t op_pc, uint32_t offset, int no
addbyte(0x75); /*JNZ +*/ addbyte(0x75); /*JNZ +*/
} }
if (not) if (not)
addbyte(5+2+2+7+5+(timing_bt ? 7 : 0)); addbyte(5+2+2+7+5+(timing_bt ? 4 : 0));
else else
addbyte(5+2+2); addbyte(5+2+2);
CALL_FUNC(CF_SET); CALL_FUNC(CF_SET);
@ -1863,40 +1863,40 @@ static void FP_FLD(int reg)
} }
addbyte(0xdd); /*FLD [ST+EAX*8]*/ addbyte(0xdd); /*FLD [ST+EAX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x83); /*AND EBX, 7*/ addbyte(0x83); /*AND EBX, 7*/
addbyte(0xe3); addbyte(0xe3);
addbyte(0x07); addbyte(0x07);
addbyte(0x8b); /*MOV EDX, [ST_i64+EAX]*/ addbyte(0x8b); /*MOV EDX, [ST_i64+EAX]*/
addbyte(0x14); addbyte(0x54);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)MM); addbyte(cpu_state_offset(MM));
addbyte(0x8b); /*MOV ECX, [ST_i64+4+EAX]*/ addbyte(0x8b); /*MOV ECX, [ST_i64+4+EAX]*/
addbyte(0x0c); addbyte(0x4c);
addbyte(0xc5); addbyte(0xc5);
addlong(((uintptr_t)MM) + 4); addbyte(cpu_state_offset(MM)+4);
addbyte(0x8a); /*MOV AL, [tag+EAX]*/ addbyte(0x8a); /*MOV AL, [tag+EAX]*/
addbyte(0x44); addbyte(0x44);
addbyte(0x05); addbyte(0x05);
addbyte(cpu_state_offset(tag[0])); addbyte(cpu_state_offset(tag[0]));
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x88); /*MOV [tag+EBX], AL*/ addbyte(0x88); /*MOV [tag+EBX], AL*/
addbyte(0x44); addbyte(0x44);
addbyte(0x1d); addbyte(0x1d);
addbyte(cpu_state_offset(tag[0])); addbyte(cpu_state_offset(tag[0]));
addbyte(0x89); /*MOV [ST_i64+EBX], EDX*/ addbyte(0x89); /*MOV [ST_i64+EBX], EDX*/
addbyte(0x14); addbyte(0x54);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)MM); addbyte(cpu_state_offset(MM));
addbyte(0x89); /*MOV [ST_i64+EBX+4], ECX*/ addbyte(0x89); /*MOV [ST_i64+EBX+4], ECX*/
addbyte(0x0c); addbyte(0x4c);
addbyte(0xdd); addbyte(0xdd);
addlong(((uintptr_t)MM) + 4); addbyte(cpu_state_offset(MM)+4);
addbyte(0x89); /*MOV [TOP], EBX*/ addbyte(0x89); /*MOV [TOP], EBX*/
addbyte(0x5d); addbyte(0x5d);
@ -1909,9 +1909,9 @@ static void FP_FST(int reg)
addbyte(0x45); addbyte(0x45);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0xdd); /*FLD [ST+EAX*8]*/ addbyte(0xdd); /*FLD [ST+EAX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x8a); /*MOV BL, [tag+EAX]*/ addbyte(0x8a); /*MOV BL, [tag+EAX]*/
addbyte(0x5c); addbyte(0x5c);
addbyte(0x05); addbyte(0x05);
@ -1928,9 +1928,9 @@ static void FP_FST(int reg)
} }
addbyte(0xdd); /*FSTP [ST+EAX*8]*/ addbyte(0xdd); /*FSTP [ST+EAX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x88); /*MOV [tag+EAX], BL*/ addbyte(0x88); /*MOV [tag+EAX], BL*/
addbyte(0x5c); addbyte(0x5c);
addbyte(0x05); addbyte(0x05);
@ -1949,24 +1949,24 @@ static void FP_FXCH(int reg)
addbyte(reg); addbyte(reg);
//#if 0 //#if 0
addbyte(0xdd); /*FLD [ST+EBX*8]*/ addbyte(0xdd); /*FLD [ST+EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x83); /*AND EAX, 7*/ addbyte(0x83); /*AND EAX, 7*/
addbyte(0xe0); addbyte(0xe0);
addbyte(0x07); addbyte(0x07);
addbyte(0xdd); /*FLD [ST+EAX*8]*/ addbyte(0xdd); /*FLD [ST+EAX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0xdd); /*FSTP [ST+EAX*8]*/ addbyte(0xdd); /*FSTP [ST+EAX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xc5); addbyte(0xc5);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
// addbyte(0xbe); /*MOVL ESI, tag*/ // addbyte(0xbe); /*MOVL ESI, tag*/
// addlong((uintptr_t)cpu_state.tag); // addlong((uintptr_t)cpu_state.tag);
addbyte(0x8a); /*MOV CL, tag[EAX]*/ addbyte(0x8a); /*MOV CL, tag[EAX]*/
@ -1986,7 +1986,7 @@ static void FP_FXCH(int reg)
addbyte(0x05); addbyte(0x05);
addbyte(cpu_state_offset(tag[0])); addbyte(cpu_state_offset(tag[0]));
addbyte(0xbe); /*MOVL ESI, ST_int64*/ addbyte(0xbe); /*MOVL ESI, ST_int64*/
addlong((uintptr_t)MM); addlong((uintptr_t)cpu_state.MM);
addbyte(0x8b); /*MOV ECX, ST_int64[EAX*8]*/ addbyte(0x8b); /*MOV ECX, ST_int64[EAX*8]*/
addbyte(0x0c); addbyte(0x0c);
addbyte(0xc6); addbyte(0xc6);
@ -2019,7 +2019,7 @@ static void FP_FXCH(int reg)
//#endif //#endif
#if 0 #if 0
addbyte(0xbe); /*MOVL ESI, ST*/ addbyte(0xbe); /*MOVL ESI, ST*/
addlong((uintptr_t)ST); addlong((uintptr_t)cpu_state.ST);
addbyte(0x8b); /*MOVL EDX, [ESI+EBX*8]*/ addbyte(0x8b); /*MOVL EDX, [ESI+EBX*8]*/
addbyte(0x14); addbyte(0x14);
@ -2081,9 +2081,9 @@ static void FP_LOAD_S()
addbyte(0x5d); addbyte(0x5d);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x0f); /*SETE [tag+EBX]*/ addbyte(0x0f); /*SETE [tag+EBX]*/
addbyte(0x94); addbyte(0x94);
addbyte(0x44); addbyte(0x44);
@ -2120,9 +2120,9 @@ static void FP_LOAD_D()
addbyte(0x5d); addbyte(0x5d);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x0f); /*SETE [tag+EBX]*/ addbyte(0x0f); /*SETE [tag+EBX]*/
addbyte(0x94); addbyte(0x94);
addbyte(0x44); addbyte(0x44);
@ -2153,9 +2153,9 @@ static void FP_LOAD_IW()
addbyte(0x5d); addbyte(0x5d);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x0f); /*SETE [tag+EBX]*/ addbyte(0x0f); /*SETE [tag+EBX]*/
addbyte(0x94); addbyte(0x94);
addbyte(0x44); addbyte(0x44);
@ -2186,9 +2186,9 @@ static void FP_LOAD_IL()
addbyte(0x5d); addbyte(0x5d);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x0f); /*SETE [tag+EBX]*/ addbyte(0x0f); /*SETE [tag+EBX]*/
addbyte(0x94); addbyte(0x94);
addbyte(0x44); addbyte(0x44);
@ -2207,29 +2207,29 @@ static void FP_LOAD_IQ()
addbyte(0xe3); addbyte(0xe3);
addbyte(7); addbyte(7);
addbyte(0x89); /*MOV [ST_i64+EBX*8], EAX*/ addbyte(0x89); /*MOV [ST_i64+EBX*8], EAX*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uint32_t)MM); addbyte(cpu_state_offset(MM));
addbyte(0x09); /*OR EAX, EDX*/ addbyte(0x09); /*OR EAX, EDX*/
addbyte(0xd0); addbyte(0xd0);
addbyte(0x89); /*MOV [ST_i64+4+EBX*8], EDX*/ addbyte(0x89); /*MOV [ST_i64+4+EBX*8], EDX*/
addbyte(0x14); addbyte(0x54);
addbyte(0xdd); addbyte(0xdd);
addlong(((uint32_t)MM) + 4); addbyte(cpu_state_offset(MM)+4);
addbyte(0x83); /*CMP EAX, 0*/ addbyte(0x83); /*CMP EAX, 0*/
addbyte(0xf8); addbyte(0xf8);
addbyte(0); addbyte(0);
addbyte(0xdf); /*FILDl [ST_i64+EBX*8]*/ addbyte(0xdf); /*FILDl [ST_i64+EBX*8]*/
addbyte(0x2c); addbyte(0x6c);
addbyte(0xdd); addbyte(0xdd);
addlong((uint32_t)MM); addbyte(cpu_state_offset(MM));
addbyte(0x0f); /*SETE AL*/ addbyte(0x0f); /*SETE AL*/
addbyte(0x94); addbyte(0x94);
addbyte(0xc0); addbyte(0xc0);
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x0c); /*OR AL, TAG_UINT64*/ addbyte(0x0c); /*OR AL, TAG_UINT64*/
addbyte(TAG_UINT64); addbyte(TAG_UINT64);
addbyte(0x89); /*MOV TOP, EBX*/ addbyte(0x89); /*MOV TOP, EBX*/
@ -2256,9 +2256,9 @@ static int FP_LOAD_REG(int reg)
addbyte(7); addbyte(7);
} }
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0xd9); /*FSTP [ESP]*/ addbyte(0xd9); /*FSTP [ESP]*/
addbyte(0x1c); addbyte(0x1c);
addbyte(0x24); addbyte(0x24);
@ -2284,9 +2284,9 @@ static void FP_LOAD_REG_D(int reg, int *host_reg1, int *host_reg2)
addbyte(7); addbyte(7);
} }
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0xdd); /*FSTP [ESP]*/ addbyte(0xdd); /*FSTP [ESP]*/
addbyte(0x1c); addbyte(0x1c);
addbyte(0x24); addbyte(0x24);
@ -2393,9 +2393,9 @@ static int FP_LOAD_REG_INT_W(int reg)
addbyte(7); addbyte(7);
} }
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x89); /*MOV [ESP+8], EAX*/ addbyte(0x89); /*MOV [ESP+8], EAX*/
addbyte(0x44); addbyte(0x44);
@ -2433,9 +2433,9 @@ static int FP_LOAD_REG_INT(int reg)
addbyte(7); addbyte(7);
} }
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x89); /*MOV [ESP+8], EAX*/ addbyte(0x89); /*MOV [ESP+8], EAX*/
addbyte(0x44); addbyte(0x44);
@ -2478,13 +2478,13 @@ static void FP_LOAD_REG_INT_Q(int reg, int *host_reg1, int *host_reg2)
has not been modified, then we can skip most of the conversion has not been modified, then we can skip most of the conversion
and just load the 64-bit integer representation directly */ and just load the 64-bit integer representation directly */
addbyte(0x8b); /*MOV ECX, [ST_i64+EBX*8]*/ addbyte(0x8b); /*MOV ECX, [ST_i64+EBX*8]*/
addbyte(0x0c); addbyte(0x4c);
addbyte(0xdd); addbyte(0xdd);
addlong((uint32_t)MM+4); addbyte(cpu_state_offset(MM)+4);
addbyte(0x8b); /*MOV EBX, [ST_i64+EBX*8]*/ addbyte(0x8b); /*MOV EBX, [ST_i64+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uint32_t)MM); addbyte(cpu_state_offset(MM));
return; return;
} }
@ -2495,24 +2495,24 @@ static void FP_LOAD_REG_INT_Q(int reg, int *host_reg1, int *host_reg2)
addbyte(cpu_state_offset(tag[0])); addbyte(cpu_state_offset(tag[0]));
addbyte(TAG_UINT64); addbyte(TAG_UINT64);
addbyte(0x74); /*JZ +*/ addbyte(0x74); /*JZ +*/
addbyte(7+7+2); addbyte(4+4+2);
addbyte(0x8b); /*MOV ECX, [ST_i64+EBX*8]*/ addbyte(0x8b); /*MOV ECX, [ST_i64+EBX*8]*/
addbyte(0x0c); addbyte(0x4c);
addbyte(0xdd); addbyte(0xdd);
addlong((uint32_t)MM+4); addbyte(cpu_state_offset(MM)+4);
addbyte(0x8b); /*MOV EBX, [ST_i64+EBX*8]*/ addbyte(0x8b); /*MOV EBX, [ST_i64+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uint32_t)MM); addbyte(cpu_state_offset(MM));
addbyte(0xeb); /*JMP done*/ addbyte(0xeb); /*JMP done*/
addbyte(7+4+3+5+2+2+4); addbyte(4+4+3+5+2+2+4);
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x89); /*MOV [ESP+8], EAX*/ addbyte(0x89); /*MOV [ESP+8], EAX*/
addbyte(0x44); addbyte(0x44);
@ -2575,9 +2575,9 @@ static void FP_OP_S(int op)
addbyte(0x04); addbyte(0x04);
addbyte(0x24); addbyte(0x24);
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/ addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/
addbyte(0x64); addbyte(0x64);
addbyte(0x1d); addbyte(0x1d);
@ -2587,9 +2587,9 @@ static void FP_OP_S(int op)
addbyte(0x04 | op); addbyte(0x04 | op);
addbyte(0x24); addbyte(0x24);
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
} }
static void FP_OP_D(int op) static void FP_OP_D(int op)
{ {
@ -2632,9 +2632,9 @@ static void FP_OP_D(int op)
addbyte(0x24); addbyte(0x24);
addbyte(0x04); addbyte(0x04);
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/ addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/
addbyte(0x64); addbyte(0x64);
addbyte(0x1d); addbyte(0x1d);
@ -2644,9 +2644,9 @@ static void FP_OP_D(int op)
addbyte(0x04 | op); addbyte(0x04 | op);
addbyte(0x24); addbyte(0x24);
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
if (((npxc >> 10) & 3) && op == FPU_ADD) if (((npxc >> 10) & 3) && op == FPU_ADD)
{ {
addbyte(0xd9); /*FLDCW [ESP+8]*/ addbyte(0xd9); /*FLDCW [ESP+8]*/
@ -2664,9 +2664,9 @@ static void FP_OP_IW(int op)
addbyte(0x04); addbyte(0x04);
addbyte(0x24); addbyte(0x24);
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/ addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/
addbyte(0x64); addbyte(0x64);
addbyte(0x1d); addbyte(0x1d);
@ -2676,9 +2676,9 @@ static void FP_OP_IW(int op)
addbyte(0x04 | op); addbyte(0x04 | op);
addbyte(0x24); addbyte(0x24);
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
} }
static void FP_OP_IL(int op) static void FP_OP_IL(int op)
{ {
@ -2689,9 +2689,9 @@ static void FP_OP_IL(int op)
addbyte(0x04); addbyte(0x04);
addbyte(0x24); addbyte(0x24);
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/ addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/
addbyte(0x64); addbyte(0x64);
addbyte(0x1d); addbyte(0x1d);
@ -2701,9 +2701,9 @@ static void FP_OP_IL(int op)
addbyte(0x04 | op); addbyte(0x04 | op);
addbyte(0x24); addbyte(0x24);
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
} }
static void FP_OP_IQ(int op) static void FP_OP_IQ(int op)
{ {
@ -2718,9 +2718,9 @@ static void FP_OP_IQ(int op)
addbyte(0x24); addbyte(0x24);
addbyte(0x04); addbyte(0x04);
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/ addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/
addbyte(0x64); addbyte(0x64);
addbyte(0x1d); addbyte(0x1d);
@ -2730,9 +2730,9 @@ static void FP_OP_IQ(int op)
addbyte(0x04 | op); addbyte(0x04 | op);
addbyte(0x24); addbyte(0x24);
addbyte(0xdd); /*FSTP [ST+EBX*8]*/ addbyte(0xdd); /*FSTP [ST+EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
} }
#define C0 (1<<8) #define C0 (1<<8)
@ -2749,9 +2749,9 @@ static void FP_COMPARE_S()
addbyte(0x04); addbyte(0x04);
addbyte(0x24); addbyte(0x24);
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x8a); /*MOV BL, [npxs+1]*/ addbyte(0x8a); /*MOV BL, [npxs+1]*/
addbyte(0x1d); addbyte(0x1d);
addlong(((uintptr_t)&npxs) + 1); addlong(((uintptr_t)&npxs) + 1);
@ -2787,9 +2787,9 @@ static void FP_COMPARE_D()
addbyte(0x24); addbyte(0x24);
addbyte(0x04); addbyte(0x04);
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x8a); /*MOV BL, [npxs+1]*/ addbyte(0x8a); /*MOV BL, [npxs+1]*/
addbyte(0x1d); addbyte(0x1d);
addlong(((uintptr_t)&npxs) + 1); addlong(((uintptr_t)&npxs) + 1);
@ -2821,9 +2821,9 @@ static void FP_COMPARE_IW()
addbyte(0x04); addbyte(0x04);
addbyte(0x24); addbyte(0x24);
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x8a); /*MOV BL, [npxs+1]*/ addbyte(0x8a); /*MOV BL, [npxs+1]*/
addbyte(0x1d); addbyte(0x1d);
addlong(((uintptr_t)&npxs) + 1); addlong(((uintptr_t)&npxs) + 1);
@ -2855,9 +2855,9 @@ static void FP_COMPARE_IL()
addbyte(0x04); addbyte(0x04);
addbyte(0x24); addbyte(0x24);
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xdd); addbyte(0xdd);
addlong((uintptr_t)ST); addbyte(cpu_state_offset(ST));
addbyte(0x8a); /*MOV BL, [npxs+1]*/ addbyte(0x8a); /*MOV BL, [npxs+1]*/
addbyte(0x1d); addbyte(0x1d);
addlong(((uintptr_t)&npxs) + 1); addlong(((uintptr_t)&npxs) + 1);
@ -2886,8 +2886,8 @@ static void FP_OP_REG(int op, int dst, int src)
addbyte(0x8b); /*MOV EAX, TOP*/ addbyte(0x8b); /*MOV EAX, TOP*/
addbyte(0x45); addbyte(0x45);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0xbe); /*MOVL ESI, ST*/ // addbyte(0xbe); /*MOVL ESI, ST*/
addlong((uintptr_t)ST); // addlong((uintptr_t)cpu_state.ST);
addbyte(0x89); /*MOV EBX, EAX*/ addbyte(0x89); /*MOV EBX, EAX*/
addbyte(0xc3); addbyte(0xc3);
if (src || dst) if (src || dst)
@ -2903,36 +2903,42 @@ static void FP_OP_REG(int op, int dst, int src)
if (src) if (src)
{ {
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/ addbyte(0x80); /*AND tag[EBX], ~TAG_UINT64*/
addbyte(0x64); addbyte(0x64);
addbyte(0x1d); addbyte(0x1d);
addbyte(cpu_state_offset(tag[0])); addbyte(cpu_state_offset(tag[0]));
addbyte(~TAG_UINT64); addbyte(~TAG_UINT64);
addbyte(0xdc); /*FADD ST[EAX*8]*/ addbyte(0xdc); /*FADD ST[EAX*8]*/
addbyte(0x04 | op); addbyte(0x44 | op);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
addbyte(0xdd); /*FSTP ST[EBX*8]*/ addbyte(0xdd); /*FSTP ST[EBX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
} }
else else
{ {
addbyte(0xdd); /*FLD [ESI+EAX*8]*/ addbyte(0xdd); /*FLD [ESI+EAX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
addbyte(0x80); /*AND tag[EAX], ~TAG_UINT64*/ addbyte(0x80); /*AND tag[EAX], ~TAG_UINT64*/
addbyte(0x64); addbyte(0x64);
addbyte(0x05); addbyte(0x05);
addbyte(cpu_state_offset(tag[0])); addbyte(cpu_state_offset(tag[0]));
addbyte(~TAG_UINT64); addbyte(~TAG_UINT64);
addbyte(0xdc); /*FADD ST[EBX*8]*/ addbyte(0xdc); /*FADD ST[EBX*8]*/
addbyte(0x04 | op); addbyte(0x44 | op);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
addbyte(0xdd); /*FSTP ST[EAX*8]*/ addbyte(0xdd); /*FSTP ST[EAX*8]*/
addbyte(0x1c); addbyte(0x5c);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
} }
} }
@ -2941,8 +2947,8 @@ static void FP_COMPARE_REG(int dst, int src)
addbyte(0x8b); /*MOV EAX, TOP*/ addbyte(0x8b); /*MOV EAX, TOP*/
addbyte(0x45); addbyte(0x45);
addbyte(cpu_state_offset(TOP)); addbyte(cpu_state_offset(TOP));
addbyte(0xbe); /*MOVL ESI, ST*/ // addbyte(0xbe); /*MOVL ESI, ST*/
addlong((uintptr_t)ST); // addlong((uintptr_t)cpu_state.ST);
addbyte(0x89); /*MOV EBX, EAX*/ addbyte(0x89); /*MOV EBX, EAX*/
addbyte(0xc3); addbyte(0xc3);
if (src || dst) if (src || dst)
@ -2967,20 +2973,24 @@ static void FP_COMPARE_REG(int dst, int src)
if (src) if (src)
{ {
addbyte(0xdd); /*FLD ST[EBX*8]*/ addbyte(0xdd); /*FLD ST[EBX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
addbyte(0xdc); /*FCOMP ST[EAX*8]*/ addbyte(0xdc); /*FCOMP ST[EAX*8]*/
addbyte(0x04 | 0x18); addbyte(0x44 | 0x18);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
} }
else else
{ {
addbyte(0xdd); /*FLD [ESI+EAX*8]*/ addbyte(0xdd); /*FLD [ESI+EAX*8]*/
addbyte(0x04); addbyte(0x44);
addbyte(0xc6); addbyte(0xc5);
addbyte(cpu_state_offset(ST));
addbyte(0xdc); /*FCOMP ST[EBX*8]*/ addbyte(0xdc); /*FCOMP ST[EBX*8]*/
addbyte(0x04 | 0x18); addbyte(0x44 | 0x18);
addbyte(0xde); addbyte(0xdd);
addbyte(cpu_state_offset(ST));
} }
addbyte(0xdf); /*FSTSW AX*/ addbyte(0xdf); /*FSTSW AX*/
@ -3133,8 +3143,8 @@ static int LOAD_MMX_D(int guest_reg)
host_reg_mapping[host_reg] = 100; host_reg_mapping[host_reg] = 100;
addbyte(0x8b); /*MOV EBX, reg*/ addbyte(0x8b); /*MOV EBX, reg*/
addbyte(0x05 | (host_reg << 3)); addbyte(0x45 | (host_reg << 3));
addlong((uint32_t)&MM[guest_reg].l[0]); addbyte(cpu_state_offset(MM[guest_reg].l[0]));
return host_reg; return host_reg;
} }
@ -3153,11 +3163,11 @@ static int LOAD_MMX_Q(int guest_reg, int *host_reg1, int *host_reg2)
} }
addbyte(0x8b); /*MOV EBX, reg*/ addbyte(0x8b); /*MOV EBX, reg*/
addbyte(0x05 | ((*host_reg1) << 3)); addbyte(0x45 | ((*host_reg1) << 3));
addlong((uint32_t)&MM[guest_reg].l[0]); addbyte(cpu_state_offset(MM[guest_reg].l[0]));
addbyte(0x8b); /*MOV ECX, reg+4*/ addbyte(0x8b); /*MOV ECX, reg+4*/
addbyte(0x05 | ((*host_reg2) << 3)); addbyte(0x45 | ((*host_reg2) << 3));
addlong((uint32_t)&MM[guest_reg].l[1]); addbyte(cpu_state_offset(MM[guest_reg].l[1]));
} }
static int LOAD_MMX_Q_MMX(int guest_reg) static int LOAD_MMX_Q_MMX(int guest_reg)
{ {
@ -3167,8 +3177,8 @@ static int LOAD_MMX_Q_MMX(int guest_reg)
addbyte(0xf3); /*MOVQ dst_reg,[reg]*/ addbyte(0xf3); /*MOVQ dst_reg,[reg]*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0x7e); addbyte(0x7e);
addbyte(0x05 | (dst_reg << 3)); addbyte(0x45 | (dst_reg << 3));
addlong((uint32_t)&MM[guest_reg].q); addbyte(cpu_state_offset(MM[guest_reg].q));
return dst_reg; return dst_reg;
} }
@ -3197,29 +3207,29 @@ static int LOAD_INT_TO_MMX(int src_reg1, int src_reg2)
static void STORE_MMX_LQ(int guest_reg, int host_reg1) static void STORE_MMX_LQ(int guest_reg, int host_reg1)
{ {
addbyte(0xC7); /*MOVL [reg],0*/ addbyte(0xC7); /*MOVL [reg],0*/
addbyte(0x05); addbyte(0x45);
addlong((uint32_t)&MM[guest_reg].l[1]); addbyte(cpu_state_offset(MM[guest_reg].l[1]));
addlong(0); addlong(0);
addbyte(0x89); /*MOVL [reg],host_reg*/ addbyte(0x89); /*MOVL [reg],host_reg*/
addbyte(0x05 | (host_reg1 << 3)); addbyte(0x45 | (host_reg1 << 3));
addlong((uint32_t)&MM[guest_reg].l[0]); addbyte(cpu_state_offset(MM[guest_reg].l[0]));
} }
static void STORE_MMX_Q(int guest_reg, int host_reg1, int host_reg2) static void STORE_MMX_Q(int guest_reg, int host_reg1, int host_reg2)
{ {
addbyte(0x89); /*MOVL [reg],host_reg*/ addbyte(0x89); /*MOVL [reg],host_reg*/
addbyte(0x05 | (host_reg1 << 3)); addbyte(0x45 | (host_reg1 << 3));
addlong((uint32_t)&MM[guest_reg].l[0]); addbyte(cpu_state_offset(MM[guest_reg].l[0]));
addbyte(0x89); /*MOVL [reg],host_reg*/ addbyte(0x89); /*MOVL [reg],host_reg*/
addbyte(0x05 | (host_reg2 << 3)); addbyte(0x45 | (host_reg2 << 3));
addlong((uint32_t)&MM[guest_reg].l[1]); addbyte(cpu_state_offset(MM[guest_reg].l[1]));
} }
static void STORE_MMX_Q_MMX(int guest_reg, int host_reg) static void STORE_MMX_Q_MMX(int guest_reg, int host_reg)
{ {
addbyte(0x66); /*MOVQ [guest_reg],host_reg*/ addbyte(0x66); /*MOVQ [guest_reg],host_reg*/
addbyte(0x0f); addbyte(0x0f);
addbyte(0xd6); addbyte(0xd6);
addbyte(0x05 | (host_reg << 3)); addbyte(0x45 | (host_reg << 3));
addlong((uint32_t)&MM[guest_reg].q); addbyte(cpu_state_offset(MM[guest_reg].q));
} }
#define MMX_x86_OP(name, opcode) \ #define MMX_x86_OP(name, opcode) \

4
src/codegen_x86-64.c
View file

@ -356,9 +356,9 @@ void codegen_block_start_recompile(codeblock_t *block)
addbyte(0x83); addbyte(0x83);
addbyte(0xEC); addbyte(0xEC);
addbyte(0x28); addbyte(0x28);
addbyte(0x48); /*MOVL EBP, &EAX*/ addbyte(0x48); /*MOVL RBP, &cpu_state*/
addbyte(0xBD); addbyte(0xBD);
addquad((uint64_t)&EAX); addquad(((uintptr_t)&cpu_state) + 128);
// pclog("New block %i for %08X %03x\n", block_current, cs+pc, block_num); // pclog("New block %i for %08X %03x\n", block_current, cs+pc, block_num);

4
src/codegen_x86.c
View file

@ -333,8 +333,8 @@ void codegen_block_start_recompile(codeblock_t *block)
addbyte(0x83); /*SUBL $16,%esp*/ addbyte(0x83); /*SUBL $16,%esp*/
addbyte(0xEC); addbyte(0xEC);
addbyte(0x10); addbyte(0x10);
addbyte(0xBD); /*MOVL EBP, &EAX*/ addbyte(0xBD); /*MOVL EBP, &cpu_state*/
addlong((uint32_t)&EAX); addlong(((uintptr_t)&cpu_state) + 128);
// pclog("New block %i for %08X %03x\n", block_current, cs+pc, block_num); // pclog("New block %i for %08X %03x\n", block_current, cs+pc, block_num);

20
src/ibm.h
View file

@ -109,6 +109,18 @@ typedef struct
int checked; /*Non-zero if selector is known to be valid*/ int checked; /*Non-zero if selector is known to be valid*/
} x86seg; } x86seg;
typedef union MMX_REG
{
uint64_t q;
int64_t sq;
uint32_t l[2];
int32_t sl[2];
uint16_t w[4];
int16_t sw[4];
uint8_t b[8];
int8_t sb[8];
} MMX_REG;
struct struct
{ {
x86reg regs[8]; x86reg regs[8];
@ -143,6 +155,12 @@ struct
int _cycles; int _cycles;
int cpu_recomp_ins; int cpu_recomp_ins;
double ST[8];
uint16_t MM_w4[8];
MMX_REG MM[8];
} cpu_state; } cpu_state;
#define cycles cpu_state._cycles #define cycles cpu_state._cycles
@ -151,7 +169,7 @@ struct
COMPILE_TIME_ASSERT(sizeof(cpu_state) <= 128); COMPILE_TIME_ASSERT(sizeof(cpu_state) <= 128);
#define cpu_state_offset(MEMBER) ((uintptr_t)&cpu_state.MEMBER - (uintptr_t)&cpu_state) #define cpu_state_offset(MEMBER) ((uintptr_t)&cpu_state.MEMBER - (uintptr_t)&cpu_state - 128)
/*x86reg regs[8];*/ /*x86reg regs[8];*/

2
src/x86_ops_mmx.h
View file

@ -6,7 +6,7 @@
#define MMX_GETSRC() \ #define MMX_GETSRC() \
if (cpu_mod == 3) \ if (cpu_mod == 3) \
{ \ { \
src = MM[cpu_rm]; \ src = cpu_state.MM[cpu_rm]; \
CLOCK_CYCLES(1); \ CLOCK_CYCLES(1); \
} \ } \
else \ else \

392
src/x86_ops_mmx_arith.h
View file

@ -6,14 +6,14 @@ static int opPADDB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] += src.b[0]; cpu_state.MM[cpu_reg].b[0] += src.b[0];
MM[cpu_reg].b[1] += src.b[1]; cpu_state.MM[cpu_reg].b[1] += src.b[1];
MM[cpu_reg].b[2] += src.b[2]; cpu_state.MM[cpu_reg].b[2] += src.b[2];
MM[cpu_reg].b[3] += src.b[3]; cpu_state.MM[cpu_reg].b[3] += src.b[3];
MM[cpu_reg].b[4] += src.b[4]; cpu_state.MM[cpu_reg].b[4] += src.b[4];
MM[cpu_reg].b[5] += src.b[5]; cpu_state.MM[cpu_reg].b[5] += src.b[5];
MM[cpu_reg].b[6] += src.b[6]; cpu_state.MM[cpu_reg].b[6] += src.b[6];
MM[cpu_reg].b[7] += src.b[7]; cpu_state.MM[cpu_reg].b[7] += src.b[7];
return 0; return 0;
} }
@ -25,14 +25,14 @@ static int opPADDB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] += src.b[0]; cpu_state.MM[cpu_reg].b[0] += src.b[0];
MM[cpu_reg].b[1] += src.b[1]; cpu_state.MM[cpu_reg].b[1] += src.b[1];
MM[cpu_reg].b[2] += src.b[2]; cpu_state.MM[cpu_reg].b[2] += src.b[2];
MM[cpu_reg].b[3] += src.b[3]; cpu_state.MM[cpu_reg].b[3] += src.b[3];
MM[cpu_reg].b[4] += src.b[4]; cpu_state.MM[cpu_reg].b[4] += src.b[4];
MM[cpu_reg].b[5] += src.b[5]; cpu_state.MM[cpu_reg].b[5] += src.b[5];
MM[cpu_reg].b[6] += src.b[6]; cpu_state.MM[cpu_reg].b[6] += src.b[6];
MM[cpu_reg].b[7] += src.b[7]; cpu_state.MM[cpu_reg].b[7] += src.b[7];
return 0; return 0;
} }
@ -45,10 +45,10 @@ static int opPADDW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] += src.w[0]; cpu_state.MM[cpu_reg].w[0] += src.w[0];
MM[cpu_reg].w[1] += src.w[1]; cpu_state.MM[cpu_reg].w[1] += src.w[1];
MM[cpu_reg].w[2] += src.w[2]; cpu_state.MM[cpu_reg].w[2] += src.w[2];
MM[cpu_reg].w[3] += src.w[3]; cpu_state.MM[cpu_reg].w[3] += src.w[3];
return 0; return 0;
} }
@ -60,10 +60,10 @@ static int opPADDW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] += src.w[0]; cpu_state.MM[cpu_reg].w[0] += src.w[0];
MM[cpu_reg].w[1] += src.w[1]; cpu_state.MM[cpu_reg].w[1] += src.w[1];
MM[cpu_reg].w[2] += src.w[2]; cpu_state.MM[cpu_reg].w[2] += src.w[2];
MM[cpu_reg].w[3] += src.w[3]; cpu_state.MM[cpu_reg].w[3] += src.w[3];
return 0; return 0;
} }
@ -76,8 +76,8 @@ static int opPADDD_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].l[0] += src.l[0]; cpu_state.MM[cpu_reg].l[0] += src.l[0];
MM[cpu_reg].l[1] += src.l[1]; cpu_state.MM[cpu_reg].l[1] += src.l[1];
return 0; return 0;
} }
@ -89,8 +89,8 @@ static int opPADDD_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].l[0] += src.l[0]; cpu_state.MM[cpu_reg].l[0] += src.l[0];
MM[cpu_reg].l[1] += src.l[1]; cpu_state.MM[cpu_reg].l[1] += src.l[1];
return 0; return 0;
} }
@ -103,14 +103,14 @@ static int opPADDSB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].sb[0] = SSATB(MM[cpu_reg].sb[0] + src.sb[0]); cpu_state.MM[cpu_reg].sb[0] = SSATB(cpu_state.MM[cpu_reg].sb[0] + src.sb[0]);
MM[cpu_reg].sb[1] = SSATB(MM[cpu_reg].sb[1] + src.sb[1]); cpu_state.MM[cpu_reg].sb[1] = SSATB(cpu_state.MM[cpu_reg].sb[1] + src.sb[1]);
MM[cpu_reg].sb[2] = SSATB(MM[cpu_reg].sb[2] + src.sb[2]); cpu_state.MM[cpu_reg].sb[2] = SSATB(cpu_state.MM[cpu_reg].sb[2] + src.sb[2]);
MM[cpu_reg].sb[3] = SSATB(MM[cpu_reg].sb[3] + src.sb[3]); cpu_state.MM[cpu_reg].sb[3] = SSATB(cpu_state.MM[cpu_reg].sb[3] + src.sb[3]);
MM[cpu_reg].sb[4] = SSATB(MM[cpu_reg].sb[4] + src.sb[4]); cpu_state.MM[cpu_reg].sb[4] = SSATB(cpu_state.MM[cpu_reg].sb[4] + src.sb[4]);
MM[cpu_reg].sb[5] = SSATB(MM[cpu_reg].sb[5] + src.sb[5]); cpu_state.MM[cpu_reg].sb[5] = SSATB(cpu_state.MM[cpu_reg].sb[5] + src.sb[5]);
MM[cpu_reg].sb[6] = SSATB(MM[cpu_reg].sb[6] + src.sb[6]); cpu_state.MM[cpu_reg].sb[6] = SSATB(cpu_state.MM[cpu_reg].sb[6] + src.sb[6]);
MM[cpu_reg].sb[7] = SSATB(MM[cpu_reg].sb[7] + src.sb[7]); cpu_state.MM[cpu_reg].sb[7] = SSATB(cpu_state.MM[cpu_reg].sb[7] + src.sb[7]);
return 0; return 0;
} }
@ -122,14 +122,14 @@ static int opPADDSB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].sb[0] = SSATB(MM[cpu_reg].sb[0] + src.sb[0]); cpu_state.MM[cpu_reg].sb[0] = SSATB(cpu_state.MM[cpu_reg].sb[0] + src.sb[0]);
MM[cpu_reg].sb[1] = SSATB(MM[cpu_reg].sb[1] + src.sb[1]); cpu_state.MM[cpu_reg].sb[1] = SSATB(cpu_state.MM[cpu_reg].sb[1] + src.sb[1]);
MM[cpu_reg].sb[2] = SSATB(MM[cpu_reg].sb[2] + src.sb[2]); cpu_state.MM[cpu_reg].sb[2] = SSATB(cpu_state.MM[cpu_reg].sb[2] + src.sb[2]);
MM[cpu_reg].sb[3] = SSATB(MM[cpu_reg].sb[3] + src.sb[3]); cpu_state.MM[cpu_reg].sb[3] = SSATB(cpu_state.MM[cpu_reg].sb[3] + src.sb[3]);
MM[cpu_reg].sb[4] = SSATB(MM[cpu_reg].sb[4] + src.sb[4]); cpu_state.MM[cpu_reg].sb[4] = SSATB(cpu_state.MM[cpu_reg].sb[4] + src.sb[4]);
MM[cpu_reg].sb[5] = SSATB(MM[cpu_reg].sb[5] + src.sb[5]); cpu_state.MM[cpu_reg].sb[5] = SSATB(cpu_state.MM[cpu_reg].sb[5] + src.sb[5]);
MM[cpu_reg].sb[6] = SSATB(MM[cpu_reg].sb[6] + src.sb[6]); cpu_state.MM[cpu_reg].sb[6] = SSATB(cpu_state.MM[cpu_reg].sb[6] + src.sb[6]);
MM[cpu_reg].sb[7] = SSATB(MM[cpu_reg].sb[7] + src.sb[7]); cpu_state.MM[cpu_reg].sb[7] = SSATB(cpu_state.MM[cpu_reg].sb[7] + src.sb[7]);
return 0; return 0;
} }
@ -142,14 +142,14 @@ static int opPADDUSB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] = USATB(MM[cpu_reg].b[0] + src.b[0]); cpu_state.MM[cpu_reg].b[0] = USATB(cpu_state.MM[cpu_reg].b[0] + src.b[0]);
MM[cpu_reg].b[1] = USATB(MM[cpu_reg].b[1] + src.b[1]); cpu_state.MM[cpu_reg].b[1] = USATB(cpu_state.MM[cpu_reg].b[1] + src.b[1]);
MM[cpu_reg].b[2] = USATB(MM[cpu_reg].b[2] + src.b[2]); cpu_state.MM[cpu_reg].b[2] = USATB(cpu_state.MM[cpu_reg].b[2] + src.b[2]);
MM[cpu_reg].b[3] = USATB(MM[cpu_reg].b[3] + src.b[3]); cpu_state.MM[cpu_reg].b[3] = USATB(cpu_state.MM[cpu_reg].b[3] + src.b[3]);
MM[cpu_reg].b[4] = USATB(MM[cpu_reg].b[4] + src.b[4]); cpu_state.MM[cpu_reg].b[4] = USATB(cpu_state.MM[cpu_reg].b[4] + src.b[4]);
MM[cpu_reg].b[5] = USATB(MM[cpu_reg].b[5] + src.b[5]); cpu_state.MM[cpu_reg].b[5] = USATB(cpu_state.MM[cpu_reg].b[5] + src.b[5]);
MM[cpu_reg].b[6] = USATB(MM[cpu_reg].b[6] + src.b[6]); cpu_state.MM[cpu_reg].b[6] = USATB(cpu_state.MM[cpu_reg].b[6] + src.b[6]);
MM[cpu_reg].b[7] = USATB(MM[cpu_reg].b[7] + src.b[7]); cpu_state.MM[cpu_reg].b[7] = USATB(cpu_state.MM[cpu_reg].b[7] + src.b[7]);
return 0; return 0;
} }
@ -161,14 +161,14 @@ static int opPADDUSB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] = USATB(MM[cpu_reg].b[0] + src.b[0]); cpu_state.MM[cpu_reg].b[0] = USATB(cpu_state.MM[cpu_reg].b[0] + src.b[0]);
MM[cpu_reg].b[1] = USATB(MM[cpu_reg].b[1] + src.b[1]); cpu_state.MM[cpu_reg].b[1] = USATB(cpu_state.MM[cpu_reg].b[1] + src.b[1]);
MM[cpu_reg].b[2] = USATB(MM[cpu_reg].b[2] + src.b[2]); cpu_state.MM[cpu_reg].b[2] = USATB(cpu_state.MM[cpu_reg].b[2] + src.b[2]);
MM[cpu_reg].b[3] = USATB(MM[cpu_reg].b[3] + src.b[3]); cpu_state.MM[cpu_reg].b[3] = USATB(cpu_state.MM[cpu_reg].b[3] + src.b[3]);
MM[cpu_reg].b[4] = USATB(MM[cpu_reg].b[4] + src.b[4]); cpu_state.MM[cpu_reg].b[4] = USATB(cpu_state.MM[cpu_reg].b[4] + src.b[4]);
MM[cpu_reg].b[5] = USATB(MM[cpu_reg].b[5] + src.b[5]); cpu_state.MM[cpu_reg].b[5] = USATB(cpu_state.MM[cpu_reg].b[5] + src.b[5]);
MM[cpu_reg].b[6] = USATB(MM[cpu_reg].b[6] + src.b[6]); cpu_state.MM[cpu_reg].b[6] = USATB(cpu_state.MM[cpu_reg].b[6] + src.b[6]);
MM[cpu_reg].b[7] = USATB(MM[cpu_reg].b[7] + src.b[7]); cpu_state.MM[cpu_reg].b[7] = USATB(cpu_state.MM[cpu_reg].b[7] + src.b[7]);
return 0; return 0;
} }
@ -181,10 +181,10 @@ static int opPADDSW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].sw[0] = SSATW(MM[cpu_reg].sw[0] + src.sw[0]); cpu_state.MM[cpu_reg].sw[0] = SSATW(cpu_state.MM[cpu_reg].sw[0] + src.sw[0]);
MM[cpu_reg].sw[1] = SSATW(MM[cpu_reg].sw[1] + src.sw[1]); cpu_state.MM[cpu_reg].sw[1] = SSATW(cpu_state.MM[cpu_reg].sw[1] + src.sw[1]);
MM[cpu_reg].sw[2] = SSATW(MM[cpu_reg].sw[2] + src.sw[2]); cpu_state.MM[cpu_reg].sw[2] = SSATW(cpu_state.MM[cpu_reg].sw[2] + src.sw[2]);
MM[cpu_reg].sw[3] = SSATW(MM[cpu_reg].sw[3] + src.sw[3]); cpu_state.MM[cpu_reg].sw[3] = SSATW(cpu_state.MM[cpu_reg].sw[3] + src.sw[3]);
return 0; return 0;
} }
@ -196,10 +196,10 @@ static int opPADDSW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].sw[0] = SSATW(MM[cpu_reg].sw[0] + src.sw[0]); cpu_state.MM[cpu_reg].sw[0] = SSATW(cpu_state.MM[cpu_reg].sw[0] + src.sw[0]);
MM[cpu_reg].sw[1] = SSATW(MM[cpu_reg].sw[1] + src.sw[1]); cpu_state.MM[cpu_reg].sw[1] = SSATW(cpu_state.MM[cpu_reg].sw[1] + src.sw[1]);
MM[cpu_reg].sw[2] = SSATW(MM[cpu_reg].sw[2] + src.sw[2]); cpu_state.MM[cpu_reg].sw[2] = SSATW(cpu_state.MM[cpu_reg].sw[2] + src.sw[2]);
MM[cpu_reg].sw[3] = SSATW(MM[cpu_reg].sw[3] + src.sw[3]); cpu_state.MM[cpu_reg].sw[3] = SSATW(cpu_state.MM[cpu_reg].sw[3] + src.sw[3]);
return 0; return 0;
} }
@ -212,10 +212,10 @@ static int opPADDUSW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] = USATW(MM[cpu_reg].w[0] + src.w[0]); cpu_state.MM[cpu_reg].w[0] = USATW(cpu_state.MM[cpu_reg].w[0] + src.w[0]);
MM[cpu_reg].w[1] = USATW(MM[cpu_reg].w[1] + src.w[1]); cpu_state.MM[cpu_reg].w[1] = USATW(cpu_state.MM[cpu_reg].w[1] + src.w[1]);
MM[cpu_reg].w[2] = USATW(MM[cpu_reg].w[2] + src.w[2]); cpu_state.MM[cpu_reg].w[2] = USATW(cpu_state.MM[cpu_reg].w[2] + src.w[2]);
MM[cpu_reg].w[3] = USATW(MM[cpu_reg].w[3] + src.w[3]); cpu_state.MM[cpu_reg].w[3] = USATW(cpu_state.MM[cpu_reg].w[3] + src.w[3]);
return 0; return 0;
} }
@ -227,10 +227,10 @@ static int opPADDUSW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] = USATW(MM[cpu_reg].w[0] + src.w[0]); cpu_state.MM[cpu_reg].w[0] = USATW(cpu_state.MM[cpu_reg].w[0] + src.w[0]);
MM[cpu_reg].w[1] = USATW(MM[cpu_reg].w[1] + src.w[1]); cpu_state.MM[cpu_reg].w[1] = USATW(cpu_state.MM[cpu_reg].w[1] + src.w[1]);
MM[cpu_reg].w[2] = USATW(MM[cpu_reg].w[2] + src.w[2]); cpu_state.MM[cpu_reg].w[2] = USATW(cpu_state.MM[cpu_reg].w[2] + src.w[2]);
MM[cpu_reg].w[3] = USATW(MM[cpu_reg].w[3] + src.w[3]); cpu_state.MM[cpu_reg].w[3] = USATW(cpu_state.MM[cpu_reg].w[3] + src.w[3]);
return 0; return 0;
} }
@ -243,15 +243,15 @@ static int opPMADDWD_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
if (MM[cpu_reg].l[0] == 0x80008000 && src.l[0] == 0x80008000) if (cpu_state.MM[cpu_reg].l[0] == 0x80008000 && src.l[0] == 0x80008000)
MM[cpu_reg].l[0] = 0x80000000; cpu_state.MM[cpu_reg].l[0] = 0x80000000;
else else
MM[cpu_reg].sl[0] = ((int32_t)MM[cpu_reg].sw[0] * (int32_t)src.sw[0]) + ((int32_t)MM[cpu_reg].sw[1] * (int32_t)src.sw[1]); cpu_state.MM[cpu_reg].sl[0] = ((int32_t)cpu_state.MM[cpu_reg].sw[0] * (int32_t)src.sw[0]) + ((int32_t)cpu_state.MM[cpu_reg].sw[1] * (int32_t)src.sw[1]);
if (MM[cpu_reg].l[1] == 0x80008000 && src.l[1] == 0x80008000) if (cpu_state.MM[cpu_reg].l[1] == 0x80008000 && src.l[1] == 0x80008000)
MM[cpu_reg].l[1] = 0x80000000; cpu_state.MM[cpu_reg].l[1] = 0x80000000;
else else
MM[cpu_reg].sl[1] = ((int32_t)MM[cpu_reg].sw[2] * (int32_t)src.sw[2]) + ((int32_t)MM[cpu_reg].sw[3] * (int32_t)src.sw[3]); cpu_state.MM[cpu_reg].sl[1] = ((int32_t)cpu_state.MM[cpu_reg].sw[2] * (int32_t)src.sw[2]) + ((int32_t)cpu_state.MM[cpu_reg].sw[3] * (int32_t)src.sw[3]);
return 0; return 0;
} }
@ -263,15 +263,15 @@ static int opPMADDWD_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
if (MM[cpu_reg].l[0] == 0x80008000 && src.l[0] == 0x80008000) if (cpu_state.MM[cpu_reg].l[0] == 0x80008000 && src.l[0] == 0x80008000)
MM[cpu_reg].l[0] = 0x80000000; cpu_state.MM[cpu_reg].l[0] = 0x80000000;
else else
MM[cpu_reg].sl[0] = ((int32_t)MM[cpu_reg].sw[0] * (int32_t)src.sw[0]) + ((int32_t)MM[cpu_reg].sw[1] * (int32_t)src.sw[1]); cpu_state.MM[cpu_reg].sl[0] = ((int32_t)cpu_state.MM[cpu_reg].sw[0] * (int32_t)src.sw[0]) + ((int32_t)cpu_state.MM[cpu_reg].sw[1] * (int32_t)src.sw[1]);
if (MM[cpu_reg].l[1] == 0x80008000 && src.l[1] == 0x80008000) if (cpu_state.MM[cpu_reg].l[1] == 0x80008000 && src.l[1] == 0x80008000)
MM[cpu_reg].l[1] = 0x80000000; cpu_state.MM[cpu_reg].l[1] = 0x80000000;
else else
MM[cpu_reg].sl[1] = ((int32_t)MM[cpu_reg].sw[2] * (int32_t)src.sw[2]) + ((int32_t)MM[cpu_reg].sw[3] * (int32_t)src.sw[3]); cpu_state.MM[cpu_reg].sl[1] = ((int32_t)cpu_state.MM[cpu_reg].sw[2] * (int32_t)src.sw[2]) + ((int32_t)cpu_state.MM[cpu_reg].sw[3] * (int32_t)src.sw[3]);
return 0; return 0;
} }
@ -284,10 +284,10 @@ static int opPMULLW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_reg].w[0] *= MM[cpu_rm].w[0]; cpu_state.MM[cpu_reg].w[0] *= cpu_state.MM[cpu_rm].w[0];
MM[cpu_reg].w[1] *= MM[cpu_rm].w[1]; cpu_state.MM[cpu_reg].w[1] *= cpu_state.MM[cpu_rm].w[1];
MM[cpu_reg].w[2] *= MM[cpu_rm].w[2]; cpu_state.MM[cpu_reg].w[2] *= cpu_state.MM[cpu_rm].w[2];
MM[cpu_reg].w[3] *= MM[cpu_rm].w[3]; cpu_state.MM[cpu_reg].w[3] *= cpu_state.MM[cpu_rm].w[3];
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
@ -296,10 +296,10 @@ static int opPMULLW_a16(uint32_t fetchdat)
src.l[0] = readmeml(easeg, cpu_state.eaaddr); src.l[0] = readmeml(easeg, cpu_state.eaaddr);
src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4); if (cpu_state.abrt) return 0; src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4); if (cpu_state.abrt) return 0;
MM[cpu_reg].w[0] *= src.w[0]; cpu_state.MM[cpu_reg].w[0] *= src.w[0];
MM[cpu_reg].w[1] *= src.w[1]; cpu_state.MM[cpu_reg].w[1] *= src.w[1];
MM[cpu_reg].w[2] *= src.w[2]; cpu_state.MM[cpu_reg].w[2] *= src.w[2];
MM[cpu_reg].w[3] *= src.w[3]; cpu_state.MM[cpu_reg].w[3] *= src.w[3];
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
return 0; return 0;
@ -311,10 +311,10 @@ static int opPMULLW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_reg].w[0] *= MM[cpu_rm].w[0]; cpu_state.MM[cpu_reg].w[0] *= cpu_state.MM[cpu_rm].w[0];
MM[cpu_reg].w[1] *= MM[cpu_rm].w[1]; cpu_state.MM[cpu_reg].w[1] *= cpu_state.MM[cpu_rm].w[1];
MM[cpu_reg].w[2] *= MM[cpu_rm].w[2]; cpu_state.MM[cpu_reg].w[2] *= cpu_state.MM[cpu_rm].w[2];
MM[cpu_reg].w[3] *= MM[cpu_rm].w[3]; cpu_state.MM[cpu_reg].w[3] *= cpu_state.MM[cpu_rm].w[3];
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
@ -323,10 +323,10 @@ static int opPMULLW_a32(uint32_t fetchdat)
src.l[0] = readmeml(easeg, cpu_state.eaaddr); src.l[0] = readmeml(easeg, cpu_state.eaaddr);
src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4); if (cpu_state.abrt) return 0; src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4); if (cpu_state.abrt) return 0;
MM[cpu_reg].w[0] *= src.w[0]; cpu_state.MM[cpu_reg].w[0] *= src.w[0];
MM[cpu_reg].w[1] *= src.w[1]; cpu_state.MM[cpu_reg].w[1] *= src.w[1];
MM[cpu_reg].w[2] *= src.w[2]; cpu_state.MM[cpu_reg].w[2] *= src.w[2];
MM[cpu_reg].w[3] *= src.w[3]; cpu_state.MM[cpu_reg].w[3] *= src.w[3];
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
return 0; return 0;
@ -339,10 +339,10 @@ static int opPMULHW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_reg].w[0] = ((int32_t)MM[cpu_reg].sw[0] * (int32_t)MM[cpu_rm].sw[0]) >> 16; cpu_state.MM[cpu_reg].w[0] = ((int32_t)cpu_state.MM[cpu_reg].sw[0] * (int32_t)cpu_state.MM[cpu_rm].sw[0]) >> 16;
MM[cpu_reg].w[1] = ((int32_t)MM[cpu_reg].sw[1] * (int32_t)MM[cpu_rm].sw[1]) >> 16; cpu_state.MM[cpu_reg].w[1] = ((int32_t)cpu_state.MM[cpu_reg].sw[1] * (int32_t)cpu_state.MM[cpu_rm].sw[1]) >> 16;
MM[cpu_reg].w[2] = ((int32_t)MM[cpu_reg].sw[2] * (int32_t)MM[cpu_rm].sw[2]) >> 16; cpu_state.MM[cpu_reg].w[2] = ((int32_t)cpu_state.MM[cpu_reg].sw[2] * (int32_t)cpu_state.MM[cpu_rm].sw[2]) >> 16;
MM[cpu_reg].w[3] = ((int32_t)MM[cpu_reg].sw[3] * (int32_t)MM[cpu_rm].sw[3]) >> 16; cpu_state.MM[cpu_reg].w[3] = ((int32_t)cpu_state.MM[cpu_reg].sw[3] * (int32_t)cpu_state.MM[cpu_rm].sw[3]) >> 16;
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
@ -351,10 +351,10 @@ static int opPMULHW_a16(uint32_t fetchdat)
src.l[0] = readmeml(easeg, cpu_state.eaaddr); src.l[0] = readmeml(easeg, cpu_state.eaaddr);
src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4); if (cpu_state.abrt) return 0; src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4); if (cpu_state.abrt) return 0;
MM[cpu_reg].w[0] = ((int32_t)MM[cpu_reg].sw[0] * (int32_t)src.sw[0]) >> 16; cpu_state.MM[cpu_reg].w[0] = ((int32_t)cpu_state.MM[cpu_reg].sw[0] * (int32_t)src.sw[0]) >> 16;
MM[cpu_reg].w[1] = ((int32_t)MM[cpu_reg].sw[1] * (int32_t)src.sw[1]) >> 16; cpu_state.MM[cpu_reg].w[1] = ((int32_t)cpu_state.MM[cpu_reg].sw[1] * (int32_t)src.sw[1]) >> 16;
MM[cpu_reg].w[2] = ((int32_t)MM[cpu_reg].sw[2] * (int32_t)src.sw[2]) >> 16; cpu_state.MM[cpu_reg].w[2] = ((int32_t)cpu_state.MM[cpu_reg].sw[2] * (int32_t)src.sw[2]) >> 16;
MM[cpu_reg].w[3] = ((int32_t)MM[cpu_reg].sw[3] * (int32_t)src.sw[3]) >> 16; cpu_state.MM[cpu_reg].w[3] = ((int32_t)cpu_state.MM[cpu_reg].sw[3] * (int32_t)src.sw[3]) >> 16;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
return 0; return 0;
@ -366,10 +366,10 @@ static int opPMULHW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_reg].w[0] = ((int32_t)MM[cpu_reg].sw[0] * (int32_t)MM[cpu_rm].sw[0]) >> 16; cpu_state.MM[cpu_reg].w[0] = ((int32_t)cpu_state.MM[cpu_reg].sw[0] * (int32_t)cpu_state.MM[cpu_rm].sw[0]) >> 16;
MM[cpu_reg].w[1] = ((int32_t)MM[cpu_reg].sw[1] * (int32_t)MM[cpu_rm].sw[1]) >> 16; cpu_state.MM[cpu_reg].w[1] = ((int32_t)cpu_state.MM[cpu_reg].sw[1] * (int32_t)cpu_state.MM[cpu_rm].sw[1]) >> 16;
MM[cpu_reg].w[2] = ((int32_t)MM[cpu_reg].sw[2] * (int32_t)MM[cpu_rm].sw[2]) >> 16; cpu_state.MM[cpu_reg].w[2] = ((int32_t)cpu_state.MM[cpu_reg].sw[2] * (int32_t)cpu_state.MM[cpu_rm].sw[2]) >> 16;
MM[cpu_reg].w[3] = ((int32_t)MM[cpu_reg].sw[3] * (int32_t)MM[cpu_rm].sw[3]) >> 16; cpu_state.MM[cpu_reg].w[3] = ((int32_t)cpu_state.MM[cpu_reg].sw[3] * (int32_t)cpu_state.MM[cpu_rm].sw[3]) >> 16;
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
@ -378,10 +378,10 @@ static int opPMULHW_a32(uint32_t fetchdat)
src.l[0] = readmeml(easeg, cpu_state.eaaddr); src.l[0] = readmeml(easeg, cpu_state.eaaddr);
src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4); if (cpu_state.abrt) return 0; src.l[1] = readmeml(easeg, cpu_state.eaaddr + 4); if (cpu_state.abrt) return 0;
MM[cpu_reg].w[0] = ((int32_t)MM[cpu_reg].sw[0] * (int32_t)src.sw[0]) >> 16; cpu_state.MM[cpu_reg].w[0] = ((int32_t)cpu_state.MM[cpu_reg].sw[0] * (int32_t)src.sw[0]) >> 16;
MM[cpu_reg].w[1] = ((int32_t)MM[cpu_reg].sw[1] * (int32_t)src.sw[1]) >> 16; cpu_state.MM[cpu_reg].w[1] = ((int32_t)cpu_state.MM[cpu_reg].sw[1] * (int32_t)src.sw[1]) >> 16;
MM[cpu_reg].w[2] = ((int32_t)MM[cpu_reg].sw[2] * (int32_t)src.sw[2]) >> 16; cpu_state.MM[cpu_reg].w[2] = ((int32_t)cpu_state.MM[cpu_reg].sw[2] * (int32_t)src.sw[2]) >> 16;
MM[cpu_reg].w[3] = ((int32_t)MM[cpu_reg].sw[3] * (int32_t)src.sw[3]) >> 16; cpu_state.MM[cpu_reg].w[3] = ((int32_t)cpu_state.MM[cpu_reg].sw[3] * (int32_t)src.sw[3]) >> 16;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
return 0; return 0;
@ -395,14 +395,14 @@ static int opPSUBB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] -= src.b[0]; cpu_state.MM[cpu_reg].b[0] -= src.b[0];
MM[cpu_reg].b[1] -= src.b[1]; cpu_state.MM[cpu_reg].b[1] -= src.b[1];
MM[cpu_reg].b[2] -= src.b[2]; cpu_state.MM[cpu_reg].b[2] -= src.b[2];
MM[cpu_reg].b[3] -= src.b[3]; cpu_state.MM[cpu_reg].b[3] -= src.b[3];
MM[cpu_reg].b[4] -= src.b[4]; cpu_state.MM[cpu_reg].b[4] -= src.b[4];
MM[cpu_reg].b[5] -= src.b[5]; cpu_state.MM[cpu_reg].b[5] -= src.b[5];
MM[cpu_reg].b[6] -= src.b[6]; cpu_state.MM[cpu_reg].b[6] -= src.b[6];
MM[cpu_reg].b[7] -= src.b[7]; cpu_state.MM[cpu_reg].b[7] -= src.b[7];
return 0; return 0;
} }
@ -414,14 +414,14 @@ static int opPSUBB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] -= src.b[0]; cpu_state.MM[cpu_reg].b[0] -= src.b[0];
MM[cpu_reg].b[1] -= src.b[1]; cpu_state.MM[cpu_reg].b[1] -= src.b[1];
MM[cpu_reg].b[2] -= src.b[2]; cpu_state.MM[cpu_reg].b[2] -= src.b[2];
MM[cpu_reg].b[3] -= src.b[3]; cpu_state.MM[cpu_reg].b[3] -= src.b[3];
MM[cpu_reg].b[4] -= src.b[4]; cpu_state.MM[cpu_reg].b[4] -= src.b[4];
MM[cpu_reg].b[5] -= src.b[5]; cpu_state.MM[cpu_reg].b[5] -= src.b[5];
MM[cpu_reg].b[6] -= src.b[6]; cpu_state.MM[cpu_reg].b[6] -= src.b[6];
MM[cpu_reg].b[7] -= src.b[7]; cpu_state.MM[cpu_reg].b[7] -= src.b[7];
return 0; return 0;
} }
@ -434,10 +434,10 @@ static int opPSUBW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] -= src.w[0]; cpu_state.MM[cpu_reg].w[0] -= src.w[0];
MM[cpu_reg].w[1] -= src.w[1]; cpu_state.MM[cpu_reg].w[1] -= src.w[1];
MM[cpu_reg].w[2] -= src.w[2]; cpu_state.MM[cpu_reg].w[2] -= src.w[2];
MM[cpu_reg].w[3] -= src.w[3]; cpu_state.MM[cpu_reg].w[3] -= src.w[3];
return 0; return 0;
} }
@ -449,10 +449,10 @@ static int opPSUBW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] -= src.w[0]; cpu_state.MM[cpu_reg].w[0] -= src.w[0];
MM[cpu_reg].w[1] -= src.w[1]; cpu_state.MM[cpu_reg].w[1] -= src.w[1];
MM[cpu_reg].w[2] -= src.w[2]; cpu_state.MM[cpu_reg].w[2] -= src.w[2];
MM[cpu_reg].w[3] -= src.w[3]; cpu_state.MM[cpu_reg].w[3] -= src.w[3];
return 0; return 0;
} }
@ -465,8 +465,8 @@ static int opPSUBD_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].l[0] -= src.l[0]; cpu_state.MM[cpu_reg].l[0] -= src.l[0];
MM[cpu_reg].l[1] -= src.l[1]; cpu_state.MM[cpu_reg].l[1] -= src.l[1];
return 0; return 0;
} }
@ -478,8 +478,8 @@ static int opPSUBD_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].l[0] -= src.l[0]; cpu_state.MM[cpu_reg].l[0] -= src.l[0];
MM[cpu_reg].l[1] -= src.l[1]; cpu_state.MM[cpu_reg].l[1] -= src.l[1];
return 0; return 0;
} }
@ -492,14 +492,14 @@ static int opPSUBSB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].sb[0] = SSATB(MM[cpu_reg].sb[0] - src.sb[0]); cpu_state.MM[cpu_reg].sb[0] = SSATB(cpu_state.MM[cpu_reg].sb[0] - src.sb[0]);
MM[cpu_reg].sb[1] = SSATB(MM[cpu_reg].sb[1] - src.sb[1]); cpu_state.MM[cpu_reg].sb[1] = SSATB(cpu_state.MM[cpu_reg].sb[1] - src.sb[1]);
MM[cpu_reg].sb[2] = SSATB(MM[cpu_reg].sb[2] - src.sb[2]); cpu_state.MM[cpu_reg].sb[2] = SSATB(cpu_state.MM[cpu_reg].sb[2] - src.sb[2]);
MM[cpu_reg].sb[3] = SSATB(MM[cpu_reg].sb[3] - src.sb[3]); cpu_state.MM[cpu_reg].sb[3] = SSATB(cpu_state.MM[cpu_reg].sb[3] - src.sb[3]);
MM[cpu_reg].sb[4] = SSATB(MM[cpu_reg].sb[4] - src.sb[4]); cpu_state.MM[cpu_reg].sb[4] = SSATB(cpu_state.MM[cpu_reg].sb[4] - src.sb[4]);
MM[cpu_reg].sb[5] = SSATB(MM[cpu_reg].sb[5] - src.sb[5]); cpu_state.MM[cpu_reg].sb[5] = SSATB(cpu_state.MM[cpu_reg].sb[5] - src.sb[5]);
MM[cpu_reg].sb[6] = SSATB(MM[cpu_reg].sb[6] - src.sb[6]); cpu_state.MM[cpu_reg].sb[6] = SSATB(cpu_state.MM[cpu_reg].sb[6] - src.sb[6]);
MM[cpu_reg].sb[7] = SSATB(MM[cpu_reg].sb[7] - src.sb[7]); cpu_state.MM[cpu_reg].sb[7] = SSATB(cpu_state.MM[cpu_reg].sb[7] - src.sb[7]);
return 0; return 0;
} }
@ -511,14 +511,14 @@ static int opPSUBSB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].sb[0] = SSATB(MM[cpu_reg].sb[0] - src.sb[0]); cpu_state.MM[cpu_reg].sb[0] = SSATB(cpu_state.MM[cpu_reg].sb[0] - src.sb[0]);
MM[cpu_reg].sb[1] = SSATB(MM[cpu_reg].sb[1] - src.sb[1]); cpu_state.MM[cpu_reg].sb[1] = SSATB(cpu_state.MM[cpu_reg].sb[1] - src.sb[1]);
MM[cpu_reg].sb[2] = SSATB(MM[cpu_reg].sb[2] - src.sb[2]); cpu_state.MM[cpu_reg].sb[2] = SSATB(cpu_state.MM[cpu_reg].sb[2] - src.sb[2]);
MM[cpu_reg].sb[3] = SSATB(MM[cpu_reg].sb[3] - src.sb[3]); cpu_state.MM[cpu_reg].sb[3] = SSATB(cpu_state.MM[cpu_reg].sb[3] - src.sb[3]);
MM[cpu_reg].sb[4] = SSATB(MM[cpu_reg].sb[4] - src.sb[4]); cpu_state.MM[cpu_reg].sb[4] = SSATB(cpu_state.MM[cpu_reg].sb[4] - src.sb[4]);
MM[cpu_reg].sb[5] = SSATB(MM[cpu_reg].sb[5] - src.sb[5]); cpu_state.MM[cpu_reg].sb[5] = SSATB(cpu_state.MM[cpu_reg].sb[5] - src.sb[5]);
MM[cpu_reg].sb[6] = SSATB(MM[cpu_reg].sb[6] - src.sb[6]); cpu_state.MM[cpu_reg].sb[6] = SSATB(cpu_state.MM[cpu_reg].sb[6] - src.sb[6]);
MM[cpu_reg].sb[7] = SSATB(MM[cpu_reg].sb[7] - src.sb[7]); cpu_state.MM[cpu_reg].sb[7] = SSATB(cpu_state.MM[cpu_reg].sb[7] - src.sb[7]);
return 0; return 0;
} }
@ -531,14 +531,14 @@ static int opPSUBUSB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] = USATB(MM[cpu_reg].b[0] - src.b[0]); cpu_state.MM[cpu_reg].b[0] = USATB(cpu_state.MM[cpu_reg].b[0] - src.b[0]);
MM[cpu_reg].b[1] = USATB(MM[cpu_reg].b[1] - src.b[1]); cpu_state.MM[cpu_reg].b[1] = USATB(cpu_state.MM[cpu_reg].b[1] - src.b[1]);
MM[cpu_reg].b[2] = USATB(MM[cpu_reg].b[2] - src.b[2]); cpu_state.MM[cpu_reg].b[2] = USATB(cpu_state.MM[cpu_reg].b[2] - src.b[2]);
MM[cpu_reg].b[3] = USATB(MM[cpu_reg].b[3] - src.b[3]); cpu_state.MM[cpu_reg].b[3] = USATB(cpu_state.MM[cpu_reg].b[3] - src.b[3]);
MM[cpu_reg].b[4] = USATB(MM[cpu_reg].b[4] - src.b[4]); cpu_state.MM[cpu_reg].b[4] = USATB(cpu_state.MM[cpu_reg].b[4] - src.b[4]);
MM[cpu_reg].b[5] = USATB(MM[cpu_reg].b[5] - src.b[5]); cpu_state.MM[cpu_reg].b[5] = USATB(cpu_state.MM[cpu_reg].b[5] - src.b[5]);
MM[cpu_reg].b[6] = USATB(MM[cpu_reg].b[6] - src.b[6]); cpu_state.MM[cpu_reg].b[6] = USATB(cpu_state.MM[cpu_reg].b[6] - src.b[6]);
MM[cpu_reg].b[7] = USATB(MM[cpu_reg].b[7] - src.b[7]); cpu_state.MM[cpu_reg].b[7] = USATB(cpu_state.MM[cpu_reg].b[7] - src.b[7]);
return 0; return 0;
} }
@ -550,14 +550,14 @@ static int opPSUBUSB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] = USATB(MM[cpu_reg].b[0] - src.b[0]); cpu_state.MM[cpu_reg].b[0] = USATB(cpu_state.MM[cpu_reg].b[0] - src.b[0]);
MM[cpu_reg].b[1] = USATB(MM[cpu_reg].b[1] - src.b[1]); cpu_state.MM[cpu_reg].b[1] = USATB(cpu_state.MM[cpu_reg].b[1] - src.b[1]);
MM[cpu_reg].b[2] = USATB(MM[cpu_reg].b[2] - src.b[2]); cpu_state.MM[cpu_reg].b[2] = USATB(cpu_state.MM[cpu_reg].b[2] - src.b[2]);
MM[cpu_reg].b[3] = USATB(MM[cpu_reg].b[3] - src.b[3]); cpu_state.MM[cpu_reg].b[3] = USATB(cpu_state.MM[cpu_reg].b[3] - src.b[3]);
MM[cpu_reg].b[4] = USATB(MM[cpu_reg].b[4] - src.b[4]); cpu_state.MM[cpu_reg].b[4] = USATB(cpu_state.MM[cpu_reg].b[4] - src.b[4]);
MM[cpu_reg].b[5] = USATB(MM[cpu_reg].b[5] - src.b[5]); cpu_state.MM[cpu_reg].b[5] = USATB(cpu_state.MM[cpu_reg].b[5] - src.b[5]);
MM[cpu_reg].b[6] = USATB(MM[cpu_reg].b[6] - src.b[6]); cpu_state.MM[cpu_reg].b[6] = USATB(cpu_state.MM[cpu_reg].b[6] - src.b[6]);
MM[cpu_reg].b[7] = USATB(MM[cpu_reg].b[7] - src.b[7]); cpu_state.MM[cpu_reg].b[7] = USATB(cpu_state.MM[cpu_reg].b[7] - src.b[7]);
return 0; return 0;
} }
@ -570,10 +570,10 @@ static int opPSUBSW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].sw[0] = SSATW(MM[cpu_reg].sw[0] - src.sw[0]); cpu_state.MM[cpu_reg].sw[0] = SSATW(cpu_state.MM[cpu_reg].sw[0] - src.sw[0]);
MM[cpu_reg].sw[1] = SSATW(MM[cpu_reg].sw[1] - src.sw[1]); cpu_state.MM[cpu_reg].sw[1] = SSATW(cpu_state.MM[cpu_reg].sw[1] - src.sw[1]);
MM[cpu_reg].sw[2] = SSATW(MM[cpu_reg].sw[2] - src.sw[2]); cpu_state.MM[cpu_reg].sw[2] = SSATW(cpu_state.MM[cpu_reg].sw[2] - src.sw[2]);
MM[cpu_reg].sw[3] = SSATW(MM[cpu_reg].sw[3] - src.sw[3]); cpu_state.MM[cpu_reg].sw[3] = SSATW(cpu_state.MM[cpu_reg].sw[3] - src.sw[3]);
return 0; return 0;
} }
@ -585,10 +585,10 @@ static int opPSUBSW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].sw[0] = SSATW(MM[cpu_reg].sw[0] - src.sw[0]); cpu_state.MM[cpu_reg].sw[0] = SSATW(cpu_state.MM[cpu_reg].sw[0] - src.sw[0]);
MM[cpu_reg].sw[1] = SSATW(MM[cpu_reg].sw[1] - src.sw[1]); cpu_state.MM[cpu_reg].sw[1] = SSATW(cpu_state.MM[cpu_reg].sw[1] - src.sw[1]);
MM[cpu_reg].sw[2] = SSATW(MM[cpu_reg].sw[2] - src.sw[2]); cpu_state.MM[cpu_reg].sw[2] = SSATW(cpu_state.MM[cpu_reg].sw[2] - src.sw[2]);
MM[cpu_reg].sw[3] = SSATW(MM[cpu_reg].sw[3] - src.sw[3]); cpu_state.MM[cpu_reg].sw[3] = SSATW(cpu_state.MM[cpu_reg].sw[3] - src.sw[3]);
return 0; return 0;
} }
@ -601,10 +601,10 @@ static int opPSUBUSW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] = USATW(MM[cpu_reg].w[0] - src.w[0]); cpu_state.MM[cpu_reg].w[0] = USATW(cpu_state.MM[cpu_reg].w[0] - src.w[0]);
MM[cpu_reg].w[1] = USATW(MM[cpu_reg].w[1] - src.w[1]); cpu_state.MM[cpu_reg].w[1] = USATW(cpu_state.MM[cpu_reg].w[1] - src.w[1]);
MM[cpu_reg].w[2] = USATW(MM[cpu_reg].w[2] - src.w[2]); cpu_state.MM[cpu_reg].w[2] = USATW(cpu_state.MM[cpu_reg].w[2] - src.w[2]);
MM[cpu_reg].w[3] = USATW(MM[cpu_reg].w[3] - src.w[3]); cpu_state.MM[cpu_reg].w[3] = USATW(cpu_state.MM[cpu_reg].w[3] - src.w[3]);
return 0; return 0;
} }
@ -616,10 +616,10 @@ static int opPSUBUSW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] = USATW(MM[cpu_reg].w[0] - src.w[0]); cpu_state.MM[cpu_reg].w[0] = USATW(cpu_state.MM[cpu_reg].w[0] - src.w[0]);
MM[cpu_reg].w[1] = USATW(MM[cpu_reg].w[1] - src.w[1]); cpu_state.MM[cpu_reg].w[1] = USATW(cpu_state.MM[cpu_reg].w[1] - src.w[1]);
MM[cpu_reg].w[2] = USATW(MM[cpu_reg].w[2] - src.w[2]); cpu_state.MM[cpu_reg].w[2] = USATW(cpu_state.MM[cpu_reg].w[2] - src.w[2]);
MM[cpu_reg].w[3] = USATW(MM[cpu_reg].w[3] - src.w[3]); cpu_state.MM[cpu_reg].w[3] = USATW(cpu_state.MM[cpu_reg].w[3] - src.w[3]);
return 0; return 0;
} }

112
src/x86_ops_mmx_cmp.h
View file

@ -7,14 +7,14 @@ static int opPCMPEQB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] = (MM[cpu_reg].b[0] == src.b[0]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[0] = (cpu_state.MM[cpu_reg].b[0] == src.b[0]) ? 0xff : 0;
MM[cpu_reg].b[1] = (MM[cpu_reg].b[1] == src.b[1]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[1] = (cpu_state.MM[cpu_reg].b[1] == src.b[1]) ? 0xff : 0;
MM[cpu_reg].b[2] = (MM[cpu_reg].b[2] == src.b[2]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[2] = (cpu_state.MM[cpu_reg].b[2] == src.b[2]) ? 0xff : 0;
MM[cpu_reg].b[3] = (MM[cpu_reg].b[3] == src.b[3]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[3] = (cpu_state.MM[cpu_reg].b[3] == src.b[3]) ? 0xff : 0;
MM[cpu_reg].b[4] = (MM[cpu_reg].b[4] == src.b[4]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[4] = (cpu_state.MM[cpu_reg].b[4] == src.b[4]) ? 0xff : 0;
MM[cpu_reg].b[5] = (MM[cpu_reg].b[5] == src.b[5]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[5] = (cpu_state.MM[cpu_reg].b[5] == src.b[5]) ? 0xff : 0;
MM[cpu_reg].b[6] = (MM[cpu_reg].b[6] == src.b[6]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[6] = (cpu_state.MM[cpu_reg].b[6] == src.b[6]) ? 0xff : 0;
MM[cpu_reg].b[7] = (MM[cpu_reg].b[7] == src.b[7]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[7] = (cpu_state.MM[cpu_reg].b[7] == src.b[7]) ? 0xff : 0;
return 0; return 0;
} }
@ -27,14 +27,14 @@ static int opPCMPEQB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] = (MM[cpu_reg].b[0] == src.b[0]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[0] = (cpu_state.MM[cpu_reg].b[0] == src.b[0]) ? 0xff : 0;
MM[cpu_reg].b[1] = (MM[cpu_reg].b[1] == src.b[1]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[1] = (cpu_state.MM[cpu_reg].b[1] == src.b[1]) ? 0xff : 0;
MM[cpu_reg].b[2] = (MM[cpu_reg].b[2] == src.b[2]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[2] = (cpu_state.MM[cpu_reg].b[2] == src.b[2]) ? 0xff : 0;
MM[cpu_reg].b[3] = (MM[cpu_reg].b[3] == src.b[3]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[3] = (cpu_state.MM[cpu_reg].b[3] == src.b[3]) ? 0xff : 0;
MM[cpu_reg].b[4] = (MM[cpu_reg].b[4] == src.b[4]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[4] = (cpu_state.MM[cpu_reg].b[4] == src.b[4]) ? 0xff : 0;
MM[cpu_reg].b[5] = (MM[cpu_reg].b[5] == src.b[5]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[5] = (cpu_state.MM[cpu_reg].b[5] == src.b[5]) ? 0xff : 0;
MM[cpu_reg].b[6] = (MM[cpu_reg].b[6] == src.b[6]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[6] = (cpu_state.MM[cpu_reg].b[6] == src.b[6]) ? 0xff : 0;
MM[cpu_reg].b[7] = (MM[cpu_reg].b[7] == src.b[7]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[7] = (cpu_state.MM[cpu_reg].b[7] == src.b[7]) ? 0xff : 0;
return 0; return 0;
} }
@ -48,14 +48,14 @@ static int opPCMPGTB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] = (MM[cpu_reg].sb[0] > src.sb[0]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[0] = (cpu_state.MM[cpu_reg].sb[0] > src.sb[0]) ? 0xff : 0;
MM[cpu_reg].b[1] = (MM[cpu_reg].sb[1] > src.sb[1]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[1] = (cpu_state.MM[cpu_reg].sb[1] > src.sb[1]) ? 0xff : 0;
MM[cpu_reg].b[2] = (MM[cpu_reg].sb[2] > src.sb[2]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[2] = (cpu_state.MM[cpu_reg].sb[2] > src.sb[2]) ? 0xff : 0;
MM[cpu_reg].b[3] = (MM[cpu_reg].sb[3] > src.sb[3]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[3] = (cpu_state.MM[cpu_reg].sb[3] > src.sb[3]) ? 0xff : 0;
MM[cpu_reg].b[4] = (MM[cpu_reg].sb[4] > src.sb[4]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[4] = (cpu_state.MM[cpu_reg].sb[4] > src.sb[4]) ? 0xff : 0;
MM[cpu_reg].b[5] = (MM[cpu_reg].sb[5] > src.sb[5]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[5] = (cpu_state.MM[cpu_reg].sb[5] > src.sb[5]) ? 0xff : 0;
MM[cpu_reg].b[6] = (MM[cpu_reg].sb[6] > src.sb[6]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[6] = (cpu_state.MM[cpu_reg].sb[6] > src.sb[6]) ? 0xff : 0;
MM[cpu_reg].b[7] = (MM[cpu_reg].sb[7] > src.sb[7]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[7] = (cpu_state.MM[cpu_reg].sb[7] > src.sb[7]) ? 0xff : 0;
return 0; return 0;
} }
@ -68,14 +68,14 @@ static int opPCMPGTB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] = (MM[cpu_reg].sb[0] > src.sb[0]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[0] = (cpu_state.MM[cpu_reg].sb[0] > src.sb[0]) ? 0xff : 0;
MM[cpu_reg].b[1] = (MM[cpu_reg].sb[1] > src.sb[1]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[1] = (cpu_state.MM[cpu_reg].sb[1] > src.sb[1]) ? 0xff : 0;
MM[cpu_reg].b[2] = (MM[cpu_reg].sb[2] > src.sb[2]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[2] = (cpu_state.MM[cpu_reg].sb[2] > src.sb[2]) ? 0xff : 0;
MM[cpu_reg].b[3] = (MM[cpu_reg].sb[3] > src.sb[3]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[3] = (cpu_state.MM[cpu_reg].sb[3] > src.sb[3]) ? 0xff : 0;
MM[cpu_reg].b[4] = (MM[cpu_reg].sb[4] > src.sb[4]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[4] = (cpu_state.MM[cpu_reg].sb[4] > src.sb[4]) ? 0xff : 0;
MM[cpu_reg].b[5] = (MM[cpu_reg].sb[5] > src.sb[5]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[5] = (cpu_state.MM[cpu_reg].sb[5] > src.sb[5]) ? 0xff : 0;
MM[cpu_reg].b[6] = (MM[cpu_reg].sb[6] > src.sb[6]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[6] = (cpu_state.MM[cpu_reg].sb[6] > src.sb[6]) ? 0xff : 0;
MM[cpu_reg].b[7] = (MM[cpu_reg].sb[7] > src.sb[7]) ? 0xff : 0; cpu_state.MM[cpu_reg].b[7] = (cpu_state.MM[cpu_reg].sb[7] > src.sb[7]) ? 0xff : 0;
return 0; return 0;
} }
@ -89,10 +89,10 @@ static int opPCMPEQW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] = (MM[cpu_reg].w[0] == src.w[0]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[0] = (cpu_state.MM[cpu_reg].w[0] == src.w[0]) ? 0xffff : 0;
MM[cpu_reg].w[1] = (MM[cpu_reg].w[1] == src.w[1]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[1] = (cpu_state.MM[cpu_reg].w[1] == src.w[1]) ? 0xffff : 0;
MM[cpu_reg].w[2] = (MM[cpu_reg].w[2] == src.w[2]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[2] = (cpu_state.MM[cpu_reg].w[2] == src.w[2]) ? 0xffff : 0;
MM[cpu_reg].w[3] = (MM[cpu_reg].w[3] == src.w[3]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[3] = (cpu_state.MM[cpu_reg].w[3] == src.w[3]) ? 0xffff : 0;
return 0; return 0;
} }
@ -105,10 +105,10 @@ static int opPCMPEQW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] = (MM[cpu_reg].w[0] == src.w[0]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[0] = (cpu_state.MM[cpu_reg].w[0] == src.w[0]) ? 0xffff : 0;
MM[cpu_reg].w[1] = (MM[cpu_reg].w[1] == src.w[1]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[1] = (cpu_state.MM[cpu_reg].w[1] == src.w[1]) ? 0xffff : 0;
MM[cpu_reg].w[2] = (MM[cpu_reg].w[2] == src.w[2]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[2] = (cpu_state.MM[cpu_reg].w[2] == src.w[2]) ? 0xffff : 0;
MM[cpu_reg].w[3] = (MM[cpu_reg].w[3] == src.w[3]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[3] = (cpu_state.MM[cpu_reg].w[3] == src.w[3]) ? 0xffff : 0;
return 0; return 0;
} }
@ -122,10 +122,10 @@ static int opPCMPGTW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] = (MM[cpu_reg].sw[0] > src.sw[0]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[0] = (cpu_state.MM[cpu_reg].sw[0] > src.sw[0]) ? 0xffff : 0;
MM[cpu_reg].w[1] = (MM[cpu_reg].sw[1] > src.sw[1]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[1] = (cpu_state.MM[cpu_reg].sw[1] > src.sw[1]) ? 0xffff : 0;
MM[cpu_reg].w[2] = (MM[cpu_reg].sw[2] > src.sw[2]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[2] = (cpu_state.MM[cpu_reg].sw[2] > src.sw[2]) ? 0xffff : 0;
MM[cpu_reg].w[3] = (MM[cpu_reg].sw[3] > src.sw[3]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[3] = (cpu_state.MM[cpu_reg].sw[3] > src.sw[3]) ? 0xffff : 0;
return 0; return 0;
} }
@ -138,10 +138,10 @@ static int opPCMPGTW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] = (MM[cpu_reg].sw[0] > src.sw[0]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[0] = (cpu_state.MM[cpu_reg].sw[0] > src.sw[0]) ? 0xffff : 0;
MM[cpu_reg].w[1] = (MM[cpu_reg].sw[1] > src.sw[1]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[1] = (cpu_state.MM[cpu_reg].sw[1] > src.sw[1]) ? 0xffff : 0;
MM[cpu_reg].w[2] = (MM[cpu_reg].sw[2] > src.sw[2]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[2] = (cpu_state.MM[cpu_reg].sw[2] > src.sw[2]) ? 0xffff : 0;
MM[cpu_reg].w[3] = (MM[cpu_reg].sw[3] > src.sw[3]) ? 0xffff : 0; cpu_state.MM[cpu_reg].w[3] = (cpu_state.MM[cpu_reg].sw[3] > src.sw[3]) ? 0xffff : 0;
return 0; return 0;
} }
@ -155,8 +155,8 @@ static int opPCMPEQD_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].l[0] = (MM[cpu_reg].l[0] == src.l[0]) ? 0xffffffff : 0; cpu_state.MM[cpu_reg].l[0] = (cpu_state.MM[cpu_reg].l[0] == src.l[0]) ? 0xffffffff : 0;
MM[cpu_reg].l[1] = (MM[cpu_reg].l[1] == src.l[1]) ? 0xffffffff : 0; cpu_state.MM[cpu_reg].l[1] = (cpu_state.MM[cpu_reg].l[1] == src.l[1]) ? 0xffffffff : 0;
return 0; return 0;
} }
@ -169,8 +169,8 @@ static int opPCMPEQD_a32(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].l[0] = (MM[cpu_reg].l[0] == src.l[0]) ? 0xffffffff : 0; cpu_state.MM[cpu_reg].l[0] = (cpu_state.MM[cpu_reg].l[0] == src.l[0]) ? 0xffffffff : 0;
MM[cpu_reg].l[1] = (MM[cpu_reg].l[1] == src.l[1]) ? 0xffffffff : 0; cpu_state.MM[cpu_reg].l[1] = (cpu_state.MM[cpu_reg].l[1] == src.l[1]) ? 0xffffffff : 0;
return 0; return 0;
} }
@ -184,8 +184,8 @@ static int opPCMPGTD_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].l[0] = (MM[cpu_reg].sl[0] > src.sl[0]) ? 0xffffffff : 0; cpu_state.MM[cpu_reg].l[0] = (cpu_state.MM[cpu_reg].sl[0] > src.sl[0]) ? 0xffffffff : 0;
MM[cpu_reg].l[1] = (MM[cpu_reg].sl[1] > src.sl[1]) ? 0xffffffff : 0; cpu_state.MM[cpu_reg].l[1] = (cpu_state.MM[cpu_reg].sl[1] > src.sl[1]) ? 0xffffffff : 0;
return 0; return 0;
} }
@ -198,8 +198,8 @@ static int opPCMPGTD_a32(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].l[0] = (MM[cpu_reg].sl[0] > src.sl[0]) ? 0xffffffff : 0; cpu_state.MM[cpu_reg].l[0] = (cpu_state.MM[cpu_reg].sl[0] > src.sl[0]) ? 0xffffffff : 0;
MM[cpu_reg].l[1] = (MM[cpu_reg].sl[1] > src.sl[1]) ? 0xffffffff : 0; cpu_state.MM[cpu_reg].l[1] = (cpu_state.MM[cpu_reg].sl[1] > src.sl[1]) ? 0xffffffff : 0;
return 0; return 0;
} }

16
src/x86_ops_mmx_logic.h
View file

@ -6,7 +6,7 @@ static int opPAND_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].q &= src.q; cpu_state.MM[cpu_reg].q &= src.q;
return 0; return 0;
} }
static int opPAND_a32(uint32_t fetchdat) static int opPAND_a32(uint32_t fetchdat)
@ -17,7 +17,7 @@ static int opPAND_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].q &= src.q; cpu_state.MM[cpu_reg].q &= src.q;
return 0; return 0;
} }
@ -29,7 +29,7 @@ static int opPANDN_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].q = ~MM[cpu_reg].q & src.q; cpu_state.MM[cpu_reg].q = ~cpu_state.MM[cpu_reg].q & src.q;
return 0; return 0;
} }
static int opPANDN_a32(uint32_t fetchdat) static int opPANDN_a32(uint32_t fetchdat)
@ -40,7 +40,7 @@ static int opPANDN_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].q = ~MM[cpu_reg].q & src.q; cpu_state.MM[cpu_reg].q = ~cpu_state.MM[cpu_reg].q & src.q;
return 0; return 0;
} }
@ -52,7 +52,7 @@ static int opPOR_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].q |= src.q; cpu_state.MM[cpu_reg].q |= src.q;
return 0; return 0;
} }
static int opPOR_a32(uint32_t fetchdat) static int opPOR_a32(uint32_t fetchdat)
@ -63,7 +63,7 @@ static int opPOR_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].q |= src.q; cpu_state.MM[cpu_reg].q |= src.q;
return 0; return 0;
} }
@ -75,7 +75,7 @@ static int opPXOR_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].q ^= src.q; cpu_state.MM[cpu_reg].q ^= src.q;
return 0; return 0;
} }
static int opPXOR_a32(uint32_t fetchdat) static int opPXOR_a32(uint32_t fetchdat)
@ -86,6 +86,6 @@ static int opPXOR_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].q ^= src.q; cpu_state.MM[cpu_reg].q ^= src.q;
return 0; return 0;
} }

40
src/x86_ops_mmx_mov.h
View file

@ -5,8 +5,8 @@ static int opMOVD_l_mm_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_reg].l[0] = cpu_state.regs[cpu_rm].l; cpu_state.MM[cpu_reg].l[0] = cpu_state.regs[cpu_rm].l;
MM[cpu_reg].l[1] = 0; cpu_state.MM[cpu_reg].l[1] = 0;
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
@ -14,8 +14,8 @@ static int opMOVD_l_mm_a16(uint32_t fetchdat)
uint32_t dst; uint32_t dst;
dst = readmeml(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 1; dst = readmeml(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 1;
MM[cpu_reg].l[0] = dst; cpu_state.MM[cpu_reg].l[0] = dst;
MM[cpu_reg].l[1] = 0; cpu_state.MM[cpu_reg].l[1] = 0;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
@ -28,8 +28,8 @@ static int opMOVD_l_mm_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_reg].l[0] = cpu_state.regs[cpu_rm].l; cpu_state.MM[cpu_reg].l[0] = cpu_state.regs[cpu_rm].l;
MM[cpu_reg].l[1] = 0; cpu_state.MM[cpu_reg].l[1] = 0;
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
@ -37,8 +37,8 @@ static int opMOVD_l_mm_a32(uint32_t fetchdat)
uint32_t dst; uint32_t dst;
dst = readmeml(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 1; dst = readmeml(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 1;
MM[cpu_reg].l[0] = dst; cpu_state.MM[cpu_reg].l[0] = dst;
MM[cpu_reg].l[1] = 0; cpu_state.MM[cpu_reg].l[1] = 0;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
@ -52,13 +52,13 @@ static int opMOVD_mm_l_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
cpu_state.regs[cpu_rm].l = MM[cpu_reg].l[0]; cpu_state.regs[cpu_rm].l = cpu_state.MM[cpu_reg].l[0];
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
{ {
CHECK_WRITE(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 3); CHECK_WRITE(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 3);
writememl(easeg, cpu_state.eaaddr, MM[cpu_reg].l[0]); if (cpu_state.abrt) return 1; writememl(easeg, cpu_state.eaaddr, cpu_state.MM[cpu_reg].l[0]); if (cpu_state.abrt) return 1;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
return 0; return 0;
@ -70,13 +70,13 @@ static int opMOVD_mm_l_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
cpu_state.regs[cpu_rm].l = MM[cpu_reg].l[0]; cpu_state.regs[cpu_rm].l = cpu_state.MM[cpu_reg].l[0];
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
{ {
CHECK_WRITE(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 3); CHECK_WRITE(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 3);
writememl(easeg, cpu_state.eaaddr, MM[cpu_reg].l[0]); if (cpu_state.abrt) return 1; writememl(easeg, cpu_state.eaaddr, cpu_state.MM[cpu_reg].l[0]); if (cpu_state.abrt) return 1;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
return 0; return 0;
@ -89,7 +89,7 @@ static int opMOVQ_q_mm_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_reg].q = MM[cpu_rm].q; cpu_state.MM[cpu_reg].q = cpu_state.MM[cpu_rm].q;
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
@ -97,7 +97,7 @@ static int opMOVQ_q_mm_a16(uint32_t fetchdat)
uint64_t dst; uint64_t dst;
dst = readmemq(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 1; dst = readmemq(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 1;
MM[cpu_reg].q = dst; cpu_state.MM[cpu_reg].q = dst;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
return 0; return 0;
@ -109,7 +109,7 @@ static int opMOVQ_q_mm_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_reg].q = MM[cpu_rm].q; cpu_state.MM[cpu_reg].q = cpu_state.MM[cpu_rm].q;
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
@ -117,7 +117,7 @@ static int opMOVQ_q_mm_a32(uint32_t fetchdat)
uint64_t dst; uint64_t dst;
dst = readmemq(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 1; dst = readmemq(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 1;
MM[cpu_reg].q = dst; cpu_state.MM[cpu_reg].q = dst;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
return 0; return 0;
@ -130,13 +130,13 @@ static int opMOVQ_mm_q_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_rm].q = MM[cpu_reg].q; cpu_state.MM[cpu_rm].q = cpu_state.MM[cpu_reg].q;
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
{ {
CHECK_WRITE(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 7); CHECK_WRITE(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 7);
writememq(easeg, cpu_state.eaaddr, MM[cpu_reg].l[0]); if (cpu_state.abrt) return 1; writememq(easeg, cpu_state.eaaddr, cpu_state.MM[cpu_reg].l[0]); if (cpu_state.abrt) return 1;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
return 0; return 0;
@ -148,13 +148,13 @@ static int opMOVQ_mm_q_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_rm].q = MM[cpu_reg].q; cpu_state.MM[cpu_rm].q = cpu_state.MM[cpu_reg].q;
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
{ {
CHECK_WRITE(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 7); CHECK_WRITE(cpu_state.ea_seg, cpu_state.eaaddr, cpu_state.eaaddr + 7);
writememq(easeg, cpu_state.eaaddr, MM[cpu_reg].q); if (cpu_state.abrt) return 1; writememq(easeg, cpu_state.eaaddr, cpu_state.MM[cpu_reg].q); if (cpu_state.abrt) return 1;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
return 0; return 0;

204
src/x86_ops_mmx_pack.h
View file

@ -5,7 +5,7 @@ static int opPUNPCKLDQ_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_reg].l[1] = MM[cpu_rm].l[0]; cpu_state.MM[cpu_reg].l[1] = cpu_state.MM[cpu_rm].l[0];
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
@ -13,7 +13,7 @@ static int opPUNPCKLDQ_a16(uint32_t fetchdat)
uint32_t src; uint32_t src;
src = readmeml(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 0; src = readmeml(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 0;
MM[cpu_reg].l[1] = src; cpu_state.MM[cpu_reg].l[1] = src;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
@ -26,7 +26,7 @@ static int opPUNPCKLDQ_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
if (cpu_mod == 3) if (cpu_mod == 3)
{ {
MM[cpu_reg].l[1] = MM[cpu_rm].l[0]; cpu_state.MM[cpu_reg].l[1] = cpu_state.MM[cpu_rm].l[0];
CLOCK_CYCLES(1); CLOCK_CYCLES(1);
} }
else else
@ -34,7 +34,7 @@ static int opPUNPCKLDQ_a32(uint32_t fetchdat)
uint32_t src; uint32_t src;
src = readmeml(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 0; src = readmeml(easeg, cpu_state.eaaddr); if (cpu_state.abrt) return 0;
MM[cpu_reg].l[1] = src; cpu_state.MM[cpu_reg].l[1] = src;
CLOCK_CYCLES(2); CLOCK_CYCLES(2);
} }
@ -49,8 +49,8 @@ static int opPUNPCKHDQ_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].l[0] = MM[cpu_reg].l[1]; cpu_state.MM[cpu_reg].l[0] = cpu_state.MM[cpu_reg].l[1];
MM[cpu_reg].l[1] = src.l[1]; cpu_state.MM[cpu_reg].l[1] = src.l[1];
return 0; return 0;
} }
@ -62,8 +62,8 @@ static int opPUNPCKHDQ_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].l[0] = MM[cpu_reg].l[1]; cpu_state.MM[cpu_reg].l[0] = cpu_state.MM[cpu_reg].l[1];
MM[cpu_reg].l[1] = src.l[1]; cpu_state.MM[cpu_reg].l[1] = src.l[1];
return 0; return 0;
} }
@ -76,14 +76,14 @@ static int opPUNPCKLBW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[7] = src.b[3]; cpu_state.MM[cpu_reg].b[7] = src.b[3];
MM[cpu_reg].b[6] = MM[cpu_reg].b[3]; cpu_state.MM[cpu_reg].b[6] = cpu_state.MM[cpu_reg].b[3];
MM[cpu_reg].b[5] = src.b[2]; cpu_state.MM[cpu_reg].b[5] = src.b[2];
MM[cpu_reg].b[4] = MM[cpu_reg].b[2]; cpu_state.MM[cpu_reg].b[4] = cpu_state.MM[cpu_reg].b[2];
MM[cpu_reg].b[3] = src.b[1]; cpu_state.MM[cpu_reg].b[3] = src.b[1];
MM[cpu_reg].b[2] = MM[cpu_reg].b[1]; cpu_state.MM[cpu_reg].b[2] = cpu_state.MM[cpu_reg].b[1];
MM[cpu_reg].b[1] = src.b[0]; cpu_state.MM[cpu_reg].b[1] = src.b[0];
MM[cpu_reg].b[0] = MM[cpu_reg].b[0]; cpu_state.MM[cpu_reg].b[0] = cpu_state.MM[cpu_reg].b[0];
return 0; return 0;
} }
@ -95,14 +95,14 @@ static int opPUNPCKLBW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[7] = src.b[3]; cpu_state.MM[cpu_reg].b[7] = src.b[3];
MM[cpu_reg].b[6] = MM[cpu_reg].b[3]; cpu_state.MM[cpu_reg].b[6] = cpu_state.MM[cpu_reg].b[3];
MM[cpu_reg].b[5] = src.b[2]; cpu_state.MM[cpu_reg].b[5] = src.b[2];
MM[cpu_reg].b[4] = MM[cpu_reg].b[2]; cpu_state.MM[cpu_reg].b[4] = cpu_state.MM[cpu_reg].b[2];
MM[cpu_reg].b[3] = src.b[1]; cpu_state.MM[cpu_reg].b[3] = src.b[1];
MM[cpu_reg].b[2] = MM[cpu_reg].b[1]; cpu_state.MM[cpu_reg].b[2] = cpu_state.MM[cpu_reg].b[1];
MM[cpu_reg].b[1] = src.b[0]; cpu_state.MM[cpu_reg].b[1] = src.b[0];
MM[cpu_reg].b[0] = MM[cpu_reg].b[0]; cpu_state.MM[cpu_reg].b[0] = cpu_state.MM[cpu_reg].b[0];
return 0; return 0;
} }
@ -115,14 +115,14 @@ static int opPUNPCKHBW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] = MM[cpu_reg].b[4]; cpu_state.MM[cpu_reg].b[0] = cpu_state.MM[cpu_reg].b[4];
MM[cpu_reg].b[1] = src.b[4]; cpu_state.MM[cpu_reg].b[1] = src.b[4];
MM[cpu_reg].b[2] = MM[cpu_reg].b[5]; cpu_state.MM[cpu_reg].b[2] = cpu_state.MM[cpu_reg].b[5];
MM[cpu_reg].b[3] = src.b[5]; cpu_state.MM[cpu_reg].b[3] = src.b[5];
MM[cpu_reg].b[4] = MM[cpu_reg].b[6]; cpu_state.MM[cpu_reg].b[4] = cpu_state.MM[cpu_reg].b[6];
MM[cpu_reg].b[5] = src.b[6]; cpu_state.MM[cpu_reg].b[5] = src.b[6];
MM[cpu_reg].b[6] = MM[cpu_reg].b[7]; cpu_state.MM[cpu_reg].b[6] = cpu_state.MM[cpu_reg].b[7];
MM[cpu_reg].b[7] = src.b[7]; cpu_state.MM[cpu_reg].b[7] = src.b[7];
return 0; return 0;
} }
@ -134,14 +134,14 @@ static int opPUNPCKHBW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].b[0] = MM[cpu_reg].b[4]; cpu_state.MM[cpu_reg].b[0] = cpu_state.MM[cpu_reg].b[4];
MM[cpu_reg].b[1] = src.b[4]; cpu_state.MM[cpu_reg].b[1] = src.b[4];
MM[cpu_reg].b[2] = MM[cpu_reg].b[5]; cpu_state.MM[cpu_reg].b[2] = cpu_state.MM[cpu_reg].b[5];
MM[cpu_reg].b[3] = src.b[5]; cpu_state.MM[cpu_reg].b[3] = src.b[5];
MM[cpu_reg].b[4] = MM[cpu_reg].b[6]; cpu_state.MM[cpu_reg].b[4] = cpu_state.MM[cpu_reg].b[6];
MM[cpu_reg].b[5] = src.b[6]; cpu_state.MM[cpu_reg].b[5] = src.b[6];
MM[cpu_reg].b[6] = MM[cpu_reg].b[7]; cpu_state.MM[cpu_reg].b[6] = cpu_state.MM[cpu_reg].b[7];
MM[cpu_reg].b[7] = src.b[7]; cpu_state.MM[cpu_reg].b[7] = src.b[7];
return 0; return 0;
} }
@ -154,10 +154,10 @@ static int opPUNPCKLWD_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[3] = src.w[1]; cpu_state.MM[cpu_reg].w[3] = src.w[1];
MM[cpu_reg].w[2] = MM[cpu_reg].w[1]; cpu_state.MM[cpu_reg].w[2] = cpu_state.MM[cpu_reg].w[1];
MM[cpu_reg].w[1] = src.w[0]; cpu_state.MM[cpu_reg].w[1] = src.w[0];
MM[cpu_reg].w[0] = MM[cpu_reg].w[0]; cpu_state.MM[cpu_reg].w[0] = cpu_state.MM[cpu_reg].w[0];
return 0; return 0;
} }
@ -169,10 +169,10 @@ static int opPUNPCKLWD_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[3] = src.w[1]; cpu_state.MM[cpu_reg].w[3] = src.w[1];
MM[cpu_reg].w[2] = MM[cpu_reg].w[1]; cpu_state.MM[cpu_reg].w[2] = cpu_state.MM[cpu_reg].w[1];
MM[cpu_reg].w[1] = src.w[0]; cpu_state.MM[cpu_reg].w[1] = src.w[0];
MM[cpu_reg].w[0] = MM[cpu_reg].w[0]; cpu_state.MM[cpu_reg].w[0] = cpu_state.MM[cpu_reg].w[0];
return 0; return 0;
} }
@ -185,10 +185,10 @@ static int opPUNPCKHWD_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] = MM[cpu_reg].w[2]; cpu_state.MM[cpu_reg].w[0] = cpu_state.MM[cpu_reg].w[2];
MM[cpu_reg].w[1] = src.w[2]; cpu_state.MM[cpu_reg].w[1] = src.w[2];
MM[cpu_reg].w[2] = MM[cpu_reg].w[3]; cpu_state.MM[cpu_reg].w[2] = cpu_state.MM[cpu_reg].w[3];
MM[cpu_reg].w[3] = src.w[3]; cpu_state.MM[cpu_reg].w[3] = src.w[3];
return 0; return 0;
} }
@ -200,10 +200,10 @@ static int opPUNPCKHWD_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
MM[cpu_reg].w[0] = MM[cpu_reg].w[2]; cpu_state.MM[cpu_reg].w[0] = cpu_state.MM[cpu_reg].w[2];
MM[cpu_reg].w[1] = src.w[2]; cpu_state.MM[cpu_reg].w[1] = src.w[2];
MM[cpu_reg].w[2] = MM[cpu_reg].w[3]; cpu_state.MM[cpu_reg].w[2] = cpu_state.MM[cpu_reg].w[3];
MM[cpu_reg].w[3] = src.w[3]; cpu_state.MM[cpu_reg].w[3] = src.w[3];
return 0; return 0;
} }
@ -215,16 +215,16 @@ static int opPACKSSWB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
dst = MM[cpu_reg]; dst = cpu_state.MM[cpu_reg];
MM[cpu_reg].sb[0] = SSATB(dst.sw[0]); cpu_state.MM[cpu_reg].sb[0] = SSATB(dst.sw[0]);
MM[cpu_reg].sb[1] = SSATB(dst.sw[1]); cpu_state.MM[cpu_reg].sb[1] = SSATB(dst.sw[1]);
MM[cpu_reg].sb[2] = SSATB(dst.sw[2]); cpu_state.MM[cpu_reg].sb[2] = SSATB(dst.sw[2]);
MM[cpu_reg].sb[3] = SSATB(dst.sw[3]); cpu_state.MM[cpu_reg].sb[3] = SSATB(dst.sw[3]);
MM[cpu_reg].sb[4] = SSATB(src.sw[0]); cpu_state.MM[cpu_reg].sb[4] = SSATB(src.sw[0]);
MM[cpu_reg].sb[5] = SSATB(src.sw[1]); cpu_state.MM[cpu_reg].sb[5] = SSATB(src.sw[1]);
MM[cpu_reg].sb[6] = SSATB(src.sw[2]); cpu_state.MM[cpu_reg].sb[6] = SSATB(src.sw[2]);
MM[cpu_reg].sb[7] = SSATB(src.sw[3]); cpu_state.MM[cpu_reg].sb[7] = SSATB(src.sw[3]);
return 0; return 0;
} }
@ -235,16 +235,16 @@ static int opPACKSSWB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
dst = MM[cpu_reg]; dst = cpu_state.MM[cpu_reg];
MM[cpu_reg].sb[0] = SSATB(dst.sw[0]); cpu_state.MM[cpu_reg].sb[0] = SSATB(dst.sw[0]);
MM[cpu_reg].sb[1] = SSATB(dst.sw[1]); cpu_state.MM[cpu_reg].sb[1] = SSATB(dst.sw[1]);
MM[cpu_reg].sb[2] = SSATB(dst.sw[2]); cpu_state.MM[cpu_reg].sb[2] = SSATB(dst.sw[2]);
MM[cpu_reg].sb[3] = SSATB(dst.sw[3]); cpu_state.MM[cpu_reg].sb[3] = SSATB(dst.sw[3]);
MM[cpu_reg].sb[4] = SSATB(src.sw[0]); cpu_state.MM[cpu_reg].sb[4] = SSATB(src.sw[0]);
MM[cpu_reg].sb[5] = SSATB(src.sw[1]); cpu_state.MM[cpu_reg].sb[5] = SSATB(src.sw[1]);
MM[cpu_reg].sb[6] = SSATB(src.sw[2]); cpu_state.MM[cpu_reg].sb[6] = SSATB(src.sw[2]);
MM[cpu_reg].sb[7] = SSATB(src.sw[3]); cpu_state.MM[cpu_reg].sb[7] = SSATB(src.sw[3]);
return 0; return 0;
} }
@ -256,16 +256,16 @@ static int opPACKUSWB_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
dst = MM[cpu_reg]; dst = cpu_state.MM[cpu_reg];
MM[cpu_reg].b[0] = USATB(dst.sw[0]); cpu_state.MM[cpu_reg].b[0] = USATB(dst.sw[0]);
MM[cpu_reg].b[1] = USATB(dst.sw[1]); cpu_state.MM[cpu_reg].b[1] = USATB(dst.sw[1]);
MM[cpu_reg].b[2] = USATB(dst.sw[2]); cpu_state.MM[cpu_reg].b[2] = USATB(dst.sw[2]);
MM[cpu_reg].b[3] = USATB(dst.sw[3]); cpu_state.MM[cpu_reg].b[3] = USATB(dst.sw[3]);
MM[cpu_reg].b[4] = USATB(src.sw[0]); cpu_state.MM[cpu_reg].b[4] = USATB(src.sw[0]);
MM[cpu_reg].b[5] = USATB(src.sw[1]); cpu_state.MM[cpu_reg].b[5] = USATB(src.sw[1]);
MM[cpu_reg].b[6] = USATB(src.sw[2]); cpu_state.MM[cpu_reg].b[6] = USATB(src.sw[2]);
MM[cpu_reg].b[7] = USATB(src.sw[3]); cpu_state.MM[cpu_reg].b[7] = USATB(src.sw[3]);
return 0; return 0;
} }
@ -276,16 +276,16 @@ static int opPACKUSWB_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
dst = MM[cpu_reg]; dst = cpu_state.MM[cpu_reg];
MM[cpu_reg].b[0] = USATB(dst.sw[0]); cpu_state.MM[cpu_reg].b[0] = USATB(dst.sw[0]);
MM[cpu_reg].b[1] = USATB(dst.sw[1]); cpu_state.MM[cpu_reg].b[1] = USATB(dst.sw[1]);
MM[cpu_reg].b[2] = USATB(dst.sw[2]); cpu_state.MM[cpu_reg].b[2] = USATB(dst.sw[2]);
MM[cpu_reg].b[3] = USATB(dst.sw[3]); cpu_state.MM[cpu_reg].b[3] = USATB(dst.sw[3]);
MM[cpu_reg].b[4] = USATB(src.sw[0]); cpu_state.MM[cpu_reg].b[4] = USATB(src.sw[0]);
MM[cpu_reg].b[5] = USATB(src.sw[1]); cpu_state.MM[cpu_reg].b[5] = USATB(src.sw[1]);
MM[cpu_reg].b[6] = USATB(src.sw[2]); cpu_state.MM[cpu_reg].b[6] = USATB(src.sw[2]);
MM[cpu_reg].b[7] = USATB(src.sw[3]); cpu_state.MM[cpu_reg].b[7] = USATB(src.sw[3]);
return 0; return 0;
} }
@ -297,12 +297,12 @@ static int opPACKSSDW_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
dst = MM[cpu_reg]; dst = cpu_state.MM[cpu_reg];
MM[cpu_reg].sw[0] = SSATW(dst.sl[0]); cpu_state.MM[cpu_reg].sw[0] = SSATW(dst.sl[0]);
MM[cpu_reg].sw[1] = SSATW(dst.sl[1]); cpu_state.MM[cpu_reg].sw[1] = SSATW(dst.sl[1]);
MM[cpu_reg].sw[2] = SSATW(src.sl[0]); cpu_state.MM[cpu_reg].sw[2] = SSATW(src.sl[0]);
MM[cpu_reg].sw[3] = SSATW(src.sl[1]); cpu_state.MM[cpu_reg].sw[3] = SSATW(src.sl[1]);
return 0; return 0;
} }
@ -313,12 +313,12 @@ static int opPACKSSDW_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
MMX_GETSRC(); MMX_GETSRC();
dst = MM[cpu_reg]; dst = cpu_state.MM[cpu_reg];
MM[cpu_reg].sw[0] = SSATW(dst.sl[0]); cpu_state.MM[cpu_reg].sw[0] = SSATW(dst.sl[0]);
MM[cpu_reg].sw[1] = SSATW(dst.sl[1]); cpu_state.MM[cpu_reg].sw[1] = SSATW(dst.sl[1]);
MM[cpu_reg].sw[2] = SSATW(src.sl[0]); cpu_state.MM[cpu_reg].sw[2] = SSATW(src.sl[0]);
MM[cpu_reg].sw[3] = SSATW(src.sl[1]); cpu_state.MM[cpu_reg].sw[3] = SSATW(src.sl[1]);
return 0; return 0;
} }

160
src/x86_ops_mmx_shift.h
View file

@ -1,7 +1,7 @@
#define MMX_GETSHIFT() \ #define MMX_GETSHIFT() \
if (cpu_mod == 3) \ if (cpu_mod == 3) \
{ \ { \
shift = MM[cpu_rm].b[0]; \ shift = cpu_state.MM[cpu_rm].b[0]; \
CLOCK_CYCLES(1); \ CLOCK_CYCLES(1); \
} \ } \
else \ else \
@ -23,32 +23,32 @@ static int opPSxxW_imm(uint32_t fetchdat)
{ {
case 0x10: /*PSRLW*/ case 0x10: /*PSRLW*/
if (shift > 15) if (shift > 15)
MM[reg].q = 0; cpu_state.MM[reg].q = 0;
else else
{ {
MM[reg].w[0] >>= shift; cpu_state.MM[reg].w[0] >>= shift;
MM[reg].w[1] >>= shift; cpu_state.MM[reg].w[1] >>= shift;
MM[reg].w[2] >>= shift; cpu_state.MM[reg].w[2] >>= shift;
MM[reg].w[3] >>= shift; cpu_state.MM[reg].w[3] >>= shift;
} }
break; break;
case 0x20: /*PSRAW*/ case 0x20: /*PSRAW*/
if (shift > 15) if (shift > 15)
shift = 15; shift = 15;
MM[reg].sw[0] >>= shift; cpu_state.MM[reg].sw[0] >>= shift;
MM[reg].sw[1] >>= shift; cpu_state.MM[reg].sw[1] >>= shift;
MM[reg].sw[2] >>= shift; cpu_state.MM[reg].sw[2] >>= shift;
MM[reg].sw[3] >>= shift; cpu_state.MM[reg].sw[3] >>= shift;
break; break;
case 0x30: /*PSLLW*/ case 0x30: /*PSLLW*/
if (shift > 15) if (shift > 15)
MM[reg].q = 0; cpu_state.MM[reg].q = 0;
else else
{ {
MM[reg].w[0] <<= shift; cpu_state.MM[reg].w[0] <<= shift;
MM[reg].w[1] <<= shift; cpu_state.MM[reg].w[1] <<= shift;
MM[reg].w[2] <<= shift; cpu_state.MM[reg].w[2] <<= shift;
MM[reg].w[3] <<= shift; cpu_state.MM[reg].w[3] <<= shift;
} }
break; break;
default: default:
@ -72,13 +72,13 @@ static int opPSLLW_a16(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 15) if (shift > 15)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
{ {
MM[cpu_reg].w[0] <<= shift; cpu_state.MM[cpu_reg].w[0] <<= shift;
MM[cpu_reg].w[1] <<= shift; cpu_state.MM[cpu_reg].w[1] <<= shift;
MM[cpu_reg].w[2] <<= shift; cpu_state.MM[cpu_reg].w[2] <<= shift;
MM[cpu_reg].w[3] <<= shift; cpu_state.MM[cpu_reg].w[3] <<= shift;
} }
return 0; return 0;
@ -93,13 +93,13 @@ static int opPSLLW_a32(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 15) if (shift > 15)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
{ {
MM[cpu_reg].w[0] <<= shift; cpu_state.MM[cpu_reg].w[0] <<= shift;
MM[cpu_reg].w[1] <<= shift; cpu_state.MM[cpu_reg].w[1] <<= shift;
MM[cpu_reg].w[2] <<= shift; cpu_state.MM[cpu_reg].w[2] <<= shift;
MM[cpu_reg].w[3] <<= shift; cpu_state.MM[cpu_reg].w[3] <<= shift;
} }
return 0; return 0;
@ -115,13 +115,13 @@ static int opPSRLW_a16(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 15) if (shift > 15)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
{ {
MM[cpu_reg].w[0] >>= shift; cpu_state.MM[cpu_reg].w[0] >>= shift;
MM[cpu_reg].w[1] >>= shift; cpu_state.MM[cpu_reg].w[1] >>= shift;
MM[cpu_reg].w[2] >>= shift; cpu_state.MM[cpu_reg].w[2] >>= shift;
MM[cpu_reg].w[3] >>= shift; cpu_state.MM[cpu_reg].w[3] >>= shift;
} }
return 0; return 0;
@ -136,13 +136,13 @@ static int opPSRLW_a32(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 15) if (shift > 15)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
{ {
MM[cpu_reg].w[0] >>= shift; cpu_state.MM[cpu_reg].w[0] >>= shift;
MM[cpu_reg].w[1] >>= shift; cpu_state.MM[cpu_reg].w[1] >>= shift;
MM[cpu_reg].w[2] >>= shift; cpu_state.MM[cpu_reg].w[2] >>= shift;
MM[cpu_reg].w[3] >>= shift; cpu_state.MM[cpu_reg].w[3] >>= shift;
} }
return 0; return 0;
@ -160,10 +160,10 @@ static int opPSRAW_a16(uint32_t fetchdat)
if (shift > 15) if (shift > 15)
shift = 15; shift = 15;
MM[cpu_reg].sw[0] >>= shift; cpu_state.MM[cpu_reg].sw[0] >>= shift;
MM[cpu_reg].sw[1] >>= shift; cpu_state.MM[cpu_reg].sw[1] >>= shift;
MM[cpu_reg].sw[2] >>= shift; cpu_state.MM[cpu_reg].sw[2] >>= shift;
MM[cpu_reg].sw[3] >>= shift; cpu_state.MM[cpu_reg].sw[3] >>= shift;
return 0; return 0;
} }
@ -179,10 +179,10 @@ static int opPSRAW_a32(uint32_t fetchdat)
if (shift > 15) if (shift > 15)
shift = 15; shift = 15;
MM[cpu_reg].sw[0] >>= shift; cpu_state.MM[cpu_reg].sw[0] >>= shift;
MM[cpu_reg].sw[1] >>= shift; cpu_state.MM[cpu_reg].sw[1] >>= shift;
MM[cpu_reg].sw[2] >>= shift; cpu_state.MM[cpu_reg].sw[2] >>= shift;
MM[cpu_reg].sw[3] >>= shift; cpu_state.MM[cpu_reg].sw[3] >>= shift;
return 0; return 0;
} }
@ -200,26 +200,26 @@ static int opPSxxD_imm(uint32_t fetchdat)
{ {
case 0x10: /*PSRLD*/ case 0x10: /*PSRLD*/
if (shift > 31) if (shift > 31)
MM[reg].q = 0; cpu_state.MM[reg].q = 0;
else else
{ {
MM[reg].l[0] >>= shift; cpu_state.MM[reg].l[0] >>= shift;
MM[reg].l[1] >>= shift; cpu_state.MM[reg].l[1] >>= shift;
} }
break; break;
case 0x20: /*PSRAD*/ case 0x20: /*PSRAD*/
if (shift > 31) if (shift > 31)
shift = 31; shift = 31;
MM[reg].sl[0] >>= shift; cpu_state.MM[reg].sl[0] >>= shift;
MM[reg].sl[1] >>= shift; cpu_state.MM[reg].sl[1] >>= shift;
break; break;
case 0x30: /*PSLLD*/ case 0x30: /*PSLLD*/
if (shift > 31) if (shift > 31)
MM[reg].q = 0; cpu_state.MM[reg].q = 0;
else else
{ {
MM[reg].l[0] <<= shift; cpu_state.MM[reg].l[0] <<= shift;
MM[reg].l[1] <<= shift; cpu_state.MM[reg].l[1] <<= shift;
} }
break; break;
default: default:
@ -243,11 +243,11 @@ static int opPSLLD_a16(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 31) if (shift > 31)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
{ {
MM[cpu_reg].l[0] <<= shift; cpu_state.MM[cpu_reg].l[0] <<= shift;
MM[cpu_reg].l[1] <<= shift; cpu_state.MM[cpu_reg].l[1] <<= shift;
} }
return 0; return 0;
@ -262,11 +262,11 @@ static int opPSLLD_a32(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 31) if (shift > 31)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
{ {
MM[cpu_reg].l[0] <<= shift; cpu_state.MM[cpu_reg].l[0] <<= shift;
MM[cpu_reg].l[1] <<= shift; cpu_state.MM[cpu_reg].l[1] <<= shift;
} }
return 0; return 0;
@ -282,11 +282,11 @@ static int opPSRLD_a16(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 31) if (shift > 31)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
{ {
MM[cpu_reg].l[0] >>= shift; cpu_state.MM[cpu_reg].l[0] >>= shift;
MM[cpu_reg].l[1] >>= shift; cpu_state.MM[cpu_reg].l[1] >>= shift;
} }
return 0; return 0;
@ -301,11 +301,11 @@ static int opPSRLD_a32(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 31) if (shift > 31)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
{ {
MM[cpu_reg].l[0] >>= shift; cpu_state.MM[cpu_reg].l[0] >>= shift;
MM[cpu_reg].l[1] >>= shift; cpu_state.MM[cpu_reg].l[1] >>= shift;
} }
return 0; return 0;
@ -323,8 +323,8 @@ static int opPSRAD_a16(uint32_t fetchdat)
if (shift > 31) if (shift > 31)
shift = 31; shift = 31;
MM[cpu_reg].sl[0] >>= shift; cpu_state.MM[cpu_reg].sl[0] >>= shift;
MM[cpu_reg].sl[1] >>= shift; cpu_state.MM[cpu_reg].sl[1] >>= shift;
return 0; return 0;
} }
@ -340,8 +340,8 @@ static int opPSRAD_a32(uint32_t fetchdat)
if (shift > 31) if (shift > 31)
shift = 31; shift = 31;
MM[cpu_reg].sl[0] >>= shift; cpu_state.MM[cpu_reg].sl[0] >>= shift;
MM[cpu_reg].sl[1] >>= shift; cpu_state.MM[cpu_reg].sl[1] >>= shift;
return 0; return 0;
} }
@ -359,20 +359,20 @@ static int opPSxxQ_imm(uint32_t fetchdat)
{ {
case 0x10: /*PSRLW*/ case 0x10: /*PSRLW*/
if (shift > 63) if (shift > 63)
MM[reg].q = 0; cpu_state.MM[reg].q = 0;
else else
MM[reg].q >>= shift; cpu_state.MM[reg].q >>= shift;
break; break;
case 0x20: /*PSRAW*/ case 0x20: /*PSRAW*/
if (shift > 63) if (shift > 63)
shift = 63; shift = 63;
MM[reg].sq >>= shift; cpu_state.MM[reg].sq >>= shift;
break; break;
case 0x30: /*PSLLW*/ case 0x30: /*PSLLW*/
if (shift > 63) if (shift > 63)
MM[reg].q = 0; cpu_state.MM[reg].q = 0;
else else
MM[reg].q <<= shift; cpu_state.MM[reg].q <<= shift;
break; break;
default: default:
pclog("Bad PSxxQ (0F 73) instruction %02X\n", op); pclog("Bad PSxxQ (0F 73) instruction %02X\n", op);
@ -395,9 +395,9 @@ static int opPSLLQ_a16(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 63) if (shift > 63)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
MM[cpu_reg].q <<= shift; cpu_state.MM[cpu_reg].q <<= shift;
return 0; return 0;
} }
@ -411,9 +411,9 @@ static int opPSLLQ_a32(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 63) if (shift > 63)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
MM[cpu_reg].q <<= shift; cpu_state.MM[cpu_reg].q <<= shift;
return 0; return 0;
} }
@ -428,9 +428,9 @@ static int opPSRLQ_a16(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 63) if (shift > 63)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
MM[cpu_reg].q >>= shift; cpu_state.MM[cpu_reg].q >>= shift;
return 0; return 0;
} }
@ -444,9 +444,9 @@ static int opPSRLQ_a32(uint32_t fetchdat)
MMX_GETSHIFT(); MMX_GETSHIFT();
if (shift > 63) if (shift > 63)
MM[cpu_reg].q = 0; cpu_state.MM[cpu_reg].q = 0;
else else
MM[cpu_reg].q >>= shift; cpu_state.MM[cpu_reg].q >>= shift;
return 0; return 0;
} }

18
src/x87.c
View file

@ -19,8 +19,6 @@
#include "x87.h" #include "x87.h"
#include "386_common.h" #include "386_common.h"
double ST[8];
MMX_REG MM[8];
uint16_t npxs,npxc; uint16_t npxs,npxc;
uint16_t x87_gettag() uint16_t x87_gettag()
@ -55,13 +53,13 @@ void x87_dumpregs()
{ {
if (cpu_state.ismmx) if (cpu_state.ismmx)
{ {
pclog("MM0=%016llX\tMM1=%016llX\tMM2=%016llX\tMM3=%016llX\n", MM[0].q, MM[1].q, MM[2].q, MM[3].q); pclog("MM0=%016llX\tMM1=%016llX\tMM2=%016llX\tMM3=%016llX\n", cpu_state.MM[0].q, cpu_state.MM[1].q, cpu_state.MM[2].q, cpu_state.MM[3].q);
pclog("MM4=%016llX\tMM5=%016llX\tMM6=%016llX\tMM7=%016llX\n", MM[4].q, MM[5].q, MM[6].q, MM[7].q); pclog("MM4=%016llX\tMM5=%016llX\tMM6=%016llX\tMM7=%016llX\n", cpu_state.MM[4].q, cpu_state.MM[5].q, cpu_state.MM[6].q, cpu_state.MM[7].q);
} }
else else
{ {
pclog("ST(0)=%f\tST(1)=%f\tST(2)=%f\tST(3)=%f\t\n",ST[cpu_state.TOP],ST[(cpu_state.TOP+1)&7],ST[(cpu_state.TOP+2)&7],ST[(cpu_state.TOP+3)&7]); pclog("ST(0)=%f\tST(1)=%f\tST(2)=%f\tST(3)=%f\t\n",cpu_state.ST[cpu_state.TOP],cpu_state.ST[(cpu_state.TOP+1)&7],cpu_state.ST[(cpu_state.TOP+2)&7],cpu_state.ST[(cpu_state.TOP+3)&7]);
pclog("ST(4)=%f\tST(5)=%f\tST(6)=%f\tST(7)=%f\t\n",ST[(cpu_state.TOP+4)&7],ST[(cpu_state.TOP+5)&7],ST[(cpu_state.TOP+6)&7],ST[(cpu_state.TOP+7)&7]); pclog("ST(4)=%f\tST(5)=%f\tST(6)=%f\tST(7)=%f\t\n",cpu_state.ST[(cpu_state.TOP+4)&7],cpu_state.ST[(cpu_state.TOP+5)&7],cpu_state.ST[(cpu_state.TOP+6)&7],cpu_state.ST[(cpu_state.TOP+7)&7]);
} }
pclog("Status = %04X Control = %04X Tag = %04X\n",npxs,npxc,x87_gettag()); pclog("Status = %04X Control = %04X Tag = %04X\n",npxs,npxc,x87_gettag());
} }
@ -70,13 +68,13 @@ void x87_print()
{ {
if (cpu_state.ismmx) if (cpu_state.ismmx)
{ {
pclog("\tMM0=%016llX\tMM1=%016llX\tMM2=%016llX\tMM3=%016llX\t", MM[0].q, MM[1].q, MM[2].q, MM[3].q); pclog("\tMM0=%016llX\tMM1=%016llX\tMM2=%016llX\tMM3=%016llX\t", cpu_state.MM[0].q, cpu_state.MM[1].q, cpu_state.MM[2].q, cpu_state.MM[3].q);
pclog("MM4=%016llX\tMM5=%016llX\tMM6=%016llX\tMM7=%016llX\n", MM[4].q, MM[5].q, MM[6].q, MM[7].q); pclog("MM4=%016llX\tMM5=%016llX\tMM6=%016llX\tMM7=%016llX\n", cpu_state.MM[4].q, cpu_state.MM[5].q, cpu_state.MM[6].q, cpu_state.MM[7].q);
} }
else else
{ {
pclog("\tST(0)=%.20f\tST(1)=%.20f\tST(2)=%f\tST(3)=%f\t",ST[cpu_state.TOP&7],ST[(cpu_state.TOP+1)&7],ST[(cpu_state.TOP+2)&7],ST[(cpu_state.TOP+3)&7]); pclog("\tST(0)=%.20f\tST(1)=%.20f\tST(2)=%f\tST(3)=%f\t",cpu_state.ST[cpu_state.TOP&7],cpu_state.ST[(cpu_state.TOP+1)&7],cpu_state.ST[(cpu_state.TOP+2)&7],cpu_state.ST[(cpu_state.TOP+3)&7]);
pclog("ST(4)=%f\tST(5)=%f\tST(6)=%f\tST(7)=%f\tTOP=%i CR=%04X SR=%04X TAG=%04X\n",ST[(cpu_state.TOP+4)&7],ST[(cpu_state.TOP+5)&7],ST[(cpu_state.TOP+6)&7],ST[(cpu_state.TOP+7)&7], cpu_state.TOP, npxc, npxs, x87_gettag()); pclog("ST(4)=%f\tST(5)=%f\tST(6)=%f\tST(7)=%f\tTOP=%i CR=%04X SR=%04X TAG=%04X\n",cpu_state.ST[(cpu_state.TOP+4)&7],cpu_state.ST[(cpu_state.TOP+5)&7],cpu_state.ST[(cpu_state.TOP+6)&7],cpu_state.ST[(cpu_state.TOP+7)&7], cpu_state.TOP, npxc, npxs, x87_gettag());
} }
} }

17
src/x87.h
View file

@ -1,21 +1,6 @@
uint32_t x87_pc_off,x87_op_off; uint32_t x87_pc_off,x87_op_off;
uint16_t x87_pc_seg,x87_op_seg; uint16_t x87_pc_seg,x87_op_seg;
extern uint16_t npxs, npxc; extern uint16_t npxs, npxc;
extern double ST[8];
typedef union MMX_REG
{
uint64_t q;
int64_t sq;
uint32_t l[2];
int32_t sl[2];
uint16_t w[5];
int16_t sw[4];
uint8_t b[8];
int8_t sb[8];
} MMX_REG;
extern MMX_REG MM[8];
static inline void x87_set_mmx(); static inline void x87_set_mmx();
static inline void x87_emms(); static inline void x87_emms();
@ -23,5 +8,5 @@ static inline void x87_emms();
uint16_t x87_gettag(); uint16_t x87_gettag();
void x87_settag(uint16_t new_tag); void x87_settag(uint16_t new_tag);
/*Hack for FPU copy. If set then ST_i64 contains the 64-bit integer loaded by FILD*/ /*Hack for FPU copy. If set then MM[].q contains the 64-bit integer loaded by FILD*/
#define TAG_UINT64 (1 << 2) #define TAG_UINT64 (1 << 2)

24
src/x87_ops.h
View file

@ -5,7 +5,7 @@
static int rounding_modes[4] = {FE_TONEAREST, FE_DOWNWARD, FE_UPWARD, FE_TOWARDZERO}; static int rounding_modes[4] = {FE_TONEAREST, FE_DOWNWARD, FE_UPWARD, FE_TOWARDZERO};
#define ST(x) ST[((cpu_state.TOP+(x))&7)] #define ST(x) cpu_state.ST[((cpu_state.TOP+(x))&7)]
#define C0 (1<<8) #define C0 (1<<8)
#define C1 (1<<9) #define C1 (1<<9)
@ -52,13 +52,13 @@ static inline void x87_checkexceptions()
static inline void x87_push(double i) static inline void x87_push(double i)
{ {
cpu_state.TOP=(cpu_state.TOP-1)&7; cpu_state.TOP=(cpu_state.TOP-1)&7;
ST[cpu_state.TOP]=i; cpu_state.ST[cpu_state.TOP] = i;
cpu_state.tag[cpu_state.TOP&7] = (i == 0.0) ? 1 : 0; cpu_state.tag[cpu_state.TOP&7] = (i == 0.0) ? 1 : 0;
} }
static inline double x87_pop() static inline double x87_pop()
{ {
double t=ST[cpu_state.TOP]; double t = cpu_state.ST[cpu_state.TOP];
cpu_state.tag[cpu_state.TOP&7] = 3; cpu_state.tag[cpu_state.TOP&7] = 3;
cpu_state.TOP=(cpu_state.TOP+1)&7; cpu_state.TOP=(cpu_state.TOP+1)&7;
return t; return t;
@ -167,12 +167,12 @@ static inline void x87_st_fsave(int reg)
if (cpu_state.tag[reg] & TAG_UINT64) if (cpu_state.tag[reg] & TAG_UINT64)
{ {
writememl(easeg, cpu_state.eaaddr, MM[reg].q & 0xffffffff); writememl(easeg, cpu_state.eaaddr, cpu_state.MM[reg].q & 0xffffffff);
writememl(easeg, cpu_state.eaaddr + 4, MM[reg].q >> 32); writememl(easeg, cpu_state.eaaddr + 4, cpu_state.MM[reg].q >> 32);
writememw(easeg, cpu_state.eaaddr + 8, 0x5555); writememw(easeg, cpu_state.eaaddr + 8, 0x5555);
} }
else else
x87_st80(ST[reg]); x87_st80(cpu_state.ST[reg]);
} }
static inline void x87_ld_frstor(int reg) static inline void x87_ld_frstor(int reg)
@ -186,19 +186,19 @@ static inline void x87_ld_frstor(int reg)
if (temp == 0x5555 && cpu_state.tag[reg] == 2) if (temp == 0x5555 && cpu_state.tag[reg] == 2)
{ {
cpu_state.tag[reg] = TAG_UINT64; cpu_state.tag[reg] = TAG_UINT64;
MM[reg].q = readmeml(easeg, cpu_state.eaaddr); cpu_state.MM[reg].q = readmeml(easeg, cpu_state.eaaddr);
MM[reg].q |= ((uint64_t)readmeml(easeg, cpu_state.eaaddr + 4) << 32); cpu_state.MM[reg].q |= ((uint64_t)readmeml(easeg, cpu_state.eaaddr + 4) << 32);
ST[reg] = (double)MM[reg].q; cpu_state.ST[reg] = (double)cpu_state.MM[reg].q;
} }
else else
ST[reg] = x87_ld80(); cpu_state.ST[reg] = x87_ld80();
} }
static inline void x87_ldmmx(MMX_REG *r) static inline void x87_ldmmx(MMX_REG *r, uint16_t *w4)
{ {
r->l[0] = readmeml(easeg, cpu_state.eaaddr); r->l[0] = readmeml(easeg, cpu_state.eaaddr);
r->l[1] = readmeml(easeg, cpu_state.eaaddr + 4); r->l[1] = readmeml(easeg, cpu_state.eaaddr + 4);
r->w[4] = readmemw(easeg, cpu_state.eaaddr + 8); *w4 = readmemw(easeg, cpu_state.eaaddr + 8);
} }
static inline void x87_stmmx(MMX_REG r) static inline void x87_stmmx(MMX_REG r)

8
src/x87_ops_loadstore.h
View file

@ -88,7 +88,7 @@ static int opFILDiq_a16(uint32_t fetchdat)
temp64 = geteaq(); if (cpu_state.abrt) return 1; temp64 = geteaq(); if (cpu_state.abrt) return 1;
if (fplog) pclog(" %f %08X %08X\n", (double)temp64, readmeml(easeg,cpu_state.eaaddr), readmeml(easeg,cpu_state.eaaddr+4)); if (fplog) pclog(" %f %08X %08X\n", (double)temp64, readmeml(easeg,cpu_state.eaaddr), readmeml(easeg,cpu_state.eaaddr+4));
x87_push((double)temp64); x87_push((double)temp64);
MM[cpu_state.TOP].q = temp64; cpu_state.MM[cpu_state.TOP].q = temp64;
cpu_state.tag[cpu_state.TOP] |= TAG_UINT64; cpu_state.tag[cpu_state.TOP] |= TAG_UINT64;
CLOCK_CYCLES(10); CLOCK_CYCLES(10);
@ -103,7 +103,7 @@ static int opFILDiq_a32(uint32_t fetchdat)
temp64 = geteaq(); if (cpu_state.abrt) return 1; temp64 = geteaq(); if (cpu_state.abrt) return 1;
if (fplog) pclog(" %f %08X %08X\n", (double)temp64, readmeml(easeg,cpu_state.eaaddr), readmeml(easeg,cpu_state.eaaddr+4)); if (fplog) pclog(" %f %08X %08X\n", (double)temp64, readmeml(easeg,cpu_state.eaaddr), readmeml(easeg,cpu_state.eaaddr+4));
x87_push((double)temp64); x87_push((double)temp64);
MM[cpu_state.TOP].q = temp64; cpu_state.MM[cpu_state.TOP].q = temp64;
cpu_state.tag[cpu_state.TOP] |= TAG_UINT64; cpu_state.tag[cpu_state.TOP] |= TAG_UINT64;
CLOCK_CYCLES(10); CLOCK_CYCLES(10);
@ -170,7 +170,7 @@ static int FISTPiq_a16(uint32_t fetchdat)
fetch_ea_16(fetchdat); fetch_ea_16(fetchdat);
if (fplog) pclog("FISTPl %08X:%08X\n", easeg, cpu_state.eaaddr); if (fplog) pclog("FISTPl %08X:%08X\n", easeg, cpu_state.eaaddr);
if (cpu_state.tag[cpu_state.TOP] & TAG_UINT64) if (cpu_state.tag[cpu_state.TOP] & TAG_UINT64)
temp64 = MM[cpu_state.TOP].q; temp64 = cpu_state.MM[cpu_state.TOP].q;
else else
temp64 = x87_fround(ST(0)); temp64 = x87_fround(ST(0));
seteaq(temp64); if (cpu_state.abrt) return 1; seteaq(temp64); if (cpu_state.abrt) return 1;
@ -185,7 +185,7 @@ static int FISTPiq_a32(uint32_t fetchdat)
fetch_ea_32(fetchdat); fetch_ea_32(fetchdat);
if (fplog) pclog("FISTPl %08X:%08X\n", easeg, cpu_state.eaaddr); if (fplog) pclog("FISTPl %08X:%08X\n", easeg, cpu_state.eaaddr);
if (cpu_state.tag[cpu_state.TOP] & TAG_UINT64) if (cpu_state.tag[cpu_state.TOP] & TAG_UINT64)
temp64 = MM[cpu_state.TOP].q; temp64 = cpu_state.MM[cpu_state.TOP].q;
else else
temp64 = x87_fround(ST(0)); temp64 = x87_fround(ST(0));
seteaq(temp64); if (cpu_state.abrt) return 1; seteaq(temp64); if (cpu_state.abrt) return 1;

96
src/x87_ops_misc.h
View file

@ -99,20 +99,20 @@ static int FSTOR()
cpu_state.eaaddr += 28; cpu_state.eaaddr += 28;
break; break;
} }
x87_ldmmx(&MM[0]); x87_ld_frstor(0); cpu_state.eaaddr += 10; x87_ldmmx(&cpu_state.MM[0], &cpu_state.MM_w4[0]); x87_ld_frstor(0); cpu_state.eaaddr += 10;
x87_ldmmx(&MM[1]); x87_ld_frstor(1); cpu_state.eaaddr += 10; x87_ldmmx(&cpu_state.MM[1], &cpu_state.MM_w4[1]); x87_ld_frstor(1); cpu_state.eaaddr += 10;
x87_ldmmx(&MM[2]); x87_ld_frstor(2); cpu_state.eaaddr += 10; x87_ldmmx(&cpu_state.MM[2], &cpu_state.MM_w4[2]); x87_ld_frstor(2); cpu_state.eaaddr += 10;
x87_ldmmx(&MM[3]); x87_ld_frstor(3); cpu_state.eaaddr += 10; x87_ldmmx(&cpu_state.MM[3], &cpu_state.MM_w4[3]); x87_ld_frstor(3); cpu_state.eaaddr += 10;
x87_ldmmx(&MM[4]); x87_ld_frstor(4); cpu_state.eaaddr += 10; x87_ldmmx(&cpu_state.MM[4], &cpu_state.MM_w4[4]); x87_ld_frstor(4); cpu_state.eaaddr += 10;
x87_ldmmx(&MM[5]); x87_ld_frstor(5); cpu_state.eaaddr += 10; x87_ldmmx(&cpu_state.MM[5], &cpu_state.MM_w4[5]); x87_ld_frstor(5); cpu_state.eaaddr += 10;
x87_ldmmx(&MM[6]); x87_ld_frstor(6); cpu_state.eaaddr += 10; x87_ldmmx(&cpu_state.MM[6], &cpu_state.MM_w4[6]); x87_ld_frstor(6); cpu_state.eaaddr += 10;
x87_ldmmx(&MM[7]); x87_ld_frstor(7); x87_ldmmx(&cpu_state.MM[7], &cpu_state.MM_w4[7]); x87_ld_frstor(7);
cpu_state.ismmx = 0; cpu_state.ismmx = 0;
/*Horrible hack, but as PCem doesn't keep the FPU stack in 80-bit precision at all times /*Horrible hack, but as PCem doesn't keep the FPU stack in 80-bit precision at all times
something like this is needed*/ something like this is needed*/
if (MM[0].w[4] == 0xffff && MM[1].w[4] == 0xffff && MM[2].w[4] == 0xffff && MM[3].w[4] == 0xffff && if (cpu_state.MM_w4[0] == 0xffff && cpu_state.MM_w4[1] == 0xffff && cpu_state.MM_w4[2] == 0xffff && cpu_state.MM_w4[3] == 0xffff &&
MM[4].w[4] == 0xffff && MM[5].w[4] == 0xffff && MM[6].w[4] == 0xffff && MM[7].w[4] == 0xffff && cpu_state.MM_w4[4] == 0xffff && cpu_state.MM_w4[5] == 0xffff && cpu_state.MM_w4[6] == 0xffff && cpu_state.MM_w4[7] == 0xffff &&
!cpu_state.TOP && !(*(uint64_t *)cpu_state.tag)) !cpu_state.TOP && !(*(uint64_t *)cpu_state.tag))
cpu_state.ismmx = 1; cpu_state.ismmx = 1;
@ -152,14 +152,14 @@ static int FSAVE()
cpu_state.eaaddr+=14; cpu_state.eaaddr+=14;
if (cpu_state.ismmx) if (cpu_state.ismmx)
{ {
x87_stmmx(MM[0]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[0]); cpu_state.eaaddr+=10;
x87_stmmx(MM[1]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[1]); cpu_state.eaaddr+=10;
x87_stmmx(MM[2]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[2]); cpu_state.eaaddr+=10;
x87_stmmx(MM[3]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[3]); cpu_state.eaaddr+=10;
x87_stmmx(MM[4]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[4]); cpu_state.eaaddr+=10;
x87_stmmx(MM[5]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[5]); cpu_state.eaaddr+=10;
x87_stmmx(MM[6]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[6]); cpu_state.eaaddr+=10;
x87_stmmx(MM[7]); x87_stmmx(cpu_state.MM[7]);
} }
else else
{ {
@ -184,14 +184,14 @@ static int FSAVE()
cpu_state.eaaddr+=14; cpu_state.eaaddr+=14;
if (cpu_state.ismmx) if (cpu_state.ismmx)
{ {
x87_stmmx(MM[0]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[0]); cpu_state.eaaddr+=10;
x87_stmmx(MM[1]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[1]); cpu_state.eaaddr+=10;
x87_stmmx(MM[2]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[2]); cpu_state.eaaddr+=10;
x87_stmmx(MM[3]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[3]); cpu_state.eaaddr+=10;
x87_stmmx(MM[4]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[4]); cpu_state.eaaddr+=10;
x87_stmmx(MM[5]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[5]); cpu_state.eaaddr+=10;
x87_stmmx(MM[6]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[6]); cpu_state.eaaddr+=10;
x87_stmmx(MM[7]); x87_stmmx(cpu_state.MM[7]);
} }
else else
{ {
@ -215,14 +215,14 @@ static int FSAVE()
cpu_state.eaaddr+=28; cpu_state.eaaddr+=28;
if (cpu_state.ismmx) if (cpu_state.ismmx)
{ {
x87_stmmx(MM[0]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[0]); cpu_state.eaaddr+=10;
x87_stmmx(MM[1]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[1]); cpu_state.eaaddr+=10;
x87_stmmx(MM[2]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[2]); cpu_state.eaaddr+=10;
x87_stmmx(MM[3]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[3]); cpu_state.eaaddr+=10;
x87_stmmx(MM[4]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[4]); cpu_state.eaaddr+=10;
x87_stmmx(MM[5]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[5]); cpu_state.eaaddr+=10;
x87_stmmx(MM[6]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[6]); cpu_state.eaaddr+=10;
x87_stmmx(MM[7]); x87_stmmx(cpu_state.MM[7]);
} }
else else
{ {
@ -247,14 +247,14 @@ static int FSAVE()
cpu_state.eaaddr+=28; cpu_state.eaaddr+=28;
if (cpu_state.ismmx) if (cpu_state.ismmx)
{ {
x87_stmmx(MM[0]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[0]); cpu_state.eaaddr+=10;
x87_stmmx(MM[1]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[1]); cpu_state.eaaddr+=10;
x87_stmmx(MM[2]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[2]); cpu_state.eaaddr+=10;
x87_stmmx(MM[3]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[3]); cpu_state.eaaddr+=10;
x87_stmmx(MM[4]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[4]); cpu_state.eaaddr+=10;
x87_stmmx(MM[5]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[5]); cpu_state.eaaddr+=10;
x87_stmmx(MM[6]); cpu_state.eaaddr+=10; x87_stmmx(cpu_state.MM[6]); cpu_state.eaaddr+=10;
x87_stmmx(MM[7]); x87_stmmx(cpu_state.MM[7]);
} }
else else
{ {
@ -316,10 +316,10 @@ static int opFLD(uint32_t fetchdat)
cpu_state.pc++; cpu_state.pc++;
if (fplog) pclog("FLD %f\n", ST(fetchdat & 7)); if (fplog) pclog("FLD %f\n", ST(fetchdat & 7));
old_tag = cpu_state.tag[(cpu_state.TOP + fetchdat) & 7]; old_tag = cpu_state.tag[(cpu_state.TOP + fetchdat) & 7];
old_i64 = MM[(cpu_state.TOP + fetchdat) & 7].q; old_i64 = cpu_state.MM[(cpu_state.TOP + fetchdat) & 7].q;
x87_push(ST(fetchdat&7)); x87_push(ST(fetchdat&7));
cpu_state.tag[cpu_state.TOP] = old_tag; cpu_state.tag[cpu_state.TOP] = old_tag;
MM[cpu_state.TOP].q = old_i64; cpu_state.MM[cpu_state.TOP].q = old_i64;
CLOCK_CYCLES(4); CLOCK_CYCLES(4);
return 0; return 0;
} }
@ -338,9 +338,9 @@ static int opFXCH(uint32_t fetchdat)
old_tag = cpu_state.tag[cpu_state.TOP]; old_tag = cpu_state.tag[cpu_state.TOP];
cpu_state.tag[cpu_state.TOP] = cpu_state.tag[(cpu_state.TOP + fetchdat) & 7]; cpu_state.tag[cpu_state.TOP] = cpu_state.tag[(cpu_state.TOP + fetchdat) & 7];
cpu_state.tag[(cpu_state.TOP + fetchdat) & 7] = old_tag; cpu_state.tag[(cpu_state.TOP + fetchdat) & 7] = old_tag;
old_i64 = MM[cpu_state.TOP].q; old_i64 = cpu_state.MM[cpu_state.TOP].q;
MM[cpu_state.TOP].q = MM[(cpu_state.TOP + fetchdat) & 7].q; cpu_state.MM[cpu_state.TOP].q = cpu_state.MM[(cpu_state.TOP + fetchdat) & 7].q;
MM[(cpu_state.TOP + fetchdat) & 7].q = old_i64; cpu_state.MM[(cpu_state.TOP + fetchdat) & 7].q = old_i64;
CLOCK_CYCLES(4); CLOCK_CYCLES(4);
return 0; return 0;
@ -808,7 +808,7 @@ static int opFSTCW_a32(uint32_t fetchdat)
if (cond_ ## condition) \ if (cond_ ## condition) \
{ \ { \
cpu_state.tag[cpu_state.TOP] = cpu_state.tag[(cpu_state.TOP + fetchdat) & 7]; \ cpu_state.tag[cpu_state.TOP] = cpu_state.tag[(cpu_state.TOP + fetchdat) & 7]; \
MM[cpu_state.TOP].q = MM[(cpu_state.TOP + fetchdat) & 7].q; \ cpu_state.MM[cpu_state.TOP].q = cpu_state.MM[(cpu_state.TOP + fetchdat) & 7].q; \
ST(0) = ST(fetchdat & 7); \ ST(0) = ST(fetchdat & 7); \
} \ } \
CLOCK_CYCLES(4); \ CLOCK_CYCLES(4); \