void JitArm::ps_rsqrte(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITPairedOff);
FALLBACK_IF(inst.Rc);
u32 b = inst.FB, d = inst.FD;
ARMReg vB0 = fpr.R0(b);
ARMReg vB1 = fpr.R1(b);
ARMReg vD0 = fpr.R0(d, false);
ARMReg vD1 = fpr.R1(d, false);
ARMReg fpscrReg = gpr.GetReg();
ARMReg V0 = D1;
ARMReg rA = gpr.GetReg();
MOVI2R(fpscrReg, (u32)&PPC_NAN);
VLDR(V0, fpscrReg, 0);
LDR(fpscrReg, R9, PPCSTATE_OFF(fpscr));
VCMP(vB0);
VMRS(_PC);
FixupBranch Less0 = B_CC(CC_LT);
VMOV(vD0, V0);
SetFPException(fpscrReg, FPSCR_VXSQRT);
FixupBranch SkipOrr0 = B();
SetJumpTarget(Less0);
SetCC(CC_EQ);
ORR(rA, rA, 1);
SetCC();
SetJumpTarget(SkipOrr0);
VCMP(vB1);
VMRS(_PC);
FixupBranch Less1 = B_CC(CC_LT);
VMOV(vD1, V0);
SetFPException(fpscrReg, FPSCR_VXSQRT);
FixupBranch SkipOrr1 = B();
SetJumpTarget(Less1);
SetCC(CC_EQ);
ORR(rA, rA, 2);
SetCC();
SetJumpTarget(SkipOrr1);
CMP(rA, 0);
FixupBranch noException = B_CC(CC_EQ);
SetFPException(fpscrReg, FPSCR_ZX);
SetJumpTarget(noException);
VCVT(S0, vB0, 0);
VCVT(S1, vB1, 0);
NEONXEmitter nemit(this);
nemit.VRSQRTE(F_32, D0, D0);
VCVT(vD0, S0, 0);
VCVT(vD1, S1, 0);
STR(fpscrReg, R9, PPCSTATE_OFF(fpscr));
gpr.Unlock(fpscrReg, rA);
}
void JitArm::fcmpu(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITFloatingPointOff)
u32 a = inst.FA, b = inst.FB;
int cr = inst.CRFD;
ARMReg vA = fpr.R0(a);
ARMReg vB = fpr.R0(b);
ARMReg fpscrReg = gpr.GetReg();
ARMReg crReg = gpr.GetReg();
Operand2 FPRFMask(0x1F, 0xA); // 0x1F000
Operand2 LessThan(0x8, 0xA); // 0x8000
Operand2 GreaterThan(0x4, 0xA); // 0x4000
Operand2 EqualTo(0x2, 0xA); // 0x2000
Operand2 NANRes(0x1, 0xA); // 0x1000
FixupBranch Done1, Done2, Done3;
LDR(fpscrReg, R9, PPCSTATE_OFF(fpscr));
BIC(fpscrReg, fpscrReg, FPRFMask);
VCMPE(vA, vB);
VMRS(_PC);
SetCC(CC_LT);
ORR(fpscrReg, fpscrReg, LessThan);
MOV(crReg, 8);
Done1 = B();
SetCC(CC_GT);
ORR(fpscrReg, fpscrReg, GreaterThan);
MOV(crReg, 4);
Done2 = B();
SetCC(CC_EQ);
ORR(fpscrReg, fpscrReg, EqualTo);
MOV(crReg, 2);
Done3 = B();
SetCC();
ORR(fpscrReg, fpscrReg, NANRes);
MOV(crReg, 1);
VCMPE(vA, vA);
VMRS(_PC);
FixupBranch NanA = B_CC(CC_NEQ);
VCMPE(vB, vB);
VMRS(_PC);
FixupBranch NanB = B_CC(CC_NEQ);
FixupBranch Done4 = B();
SetJumpTarget(NanA);
SetJumpTarget(NanB);
SetFPException(fpscrReg, FPSCR_VXSNAN);
SetJumpTarget(Done1);
SetJumpTarget(Done2);
SetJumpTarget(Done3);
SetJumpTarget(Done4);
STRB(crReg, R9, PPCSTATE_OFF(cr_fast) + cr);
STR(fpscrReg, R9, PPCSTATE_OFF(fpscr));
gpr.Unlock(fpscrReg, crReg);
}
void JitArm::lwz(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(LoadStore)
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
ARMReg RD = gpr.R(inst.RD);
LDR(rA, R9, PPCSTATE_OFF(Exceptions));
CMP(rA, EXCEPTION_DSI);
FixupBranch DoNotLoad = B_CC(CC_EQ);
{
if (inst.RA)
{
MOVI2R(rB, inst.SIMM_16);
ARMReg RA = gpr.R(inst.RA);
ADD(rB, rB, RA);
}
else
MOVI2R(rB, (u32)inst.SIMM_16);
MOVI2R(rA, (u32)&Memory::Read_U32);
PUSH(4, R0, R1, R2, R3);
MOV(R0, rB);
BL(rA);
MOV(rA, R0);
POP(4, R0, R1, R2, R3);
MOV(RD, rA);
gpr.Unlock(rA, rB);
}
SetJumpTarget(DoNotLoad);
if (SConfig::GetInstance().m_LocalCoreStartupParameter.bSkipIdle &&
(inst.hex & 0xFFFF0000) == 0x800D0000 &&
(Memory::ReadUnchecked_U32(js.compilerPC + 4) == 0x28000000 ||
(SConfig::GetInstance().m_LocalCoreStartupParameter.bWii && Memory::ReadUnchecked_U32(js.compilerPC + 4) == 0x2C000000)) &&
Memory::ReadUnchecked_U32(js.compilerPC + 8) == 0x4182fff8)
{
gpr.Flush();
fpr.Flush();
// if it's still 0, we can wait until the next event
TST(RD, RD);
FixupBranch noIdle = B_CC(CC_NEQ);
rA = gpr.GetReg();
MOVI2R(rA, (u32)&PowerPC::OnIdle);
MOVI2R(R0, PowerPC::ppcState.gpr[inst.RA] + (s32)(s16)inst.SIMM_16);
BL(rA);
gpr.Unlock(rA);
WriteExceptionExit();
SetJumpTarget(noIdle);
//js.compilerPC += 8;
return;
}
}
void JitArm::stbu(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(LoadStore)
ARMReg RA = gpr.R(inst.RA);
ARMReg RS = gpr.R(inst.RS);
ARMReg ValueReg = gpr.GetReg();
ARMReg Addr = gpr.GetReg();
ARMReg Function = gpr.GetReg();
MOVI2R(Addr, inst.SIMM_16);
ADD(Addr, Addr, RA);
// Check for DSI exception prior to writing back address
LDR(Function, R9, PPCSTATE_OFF(Exceptions));
CMP(Function, EXCEPTION_DSI);
FixupBranch DoNotWrite = B_CC(CC_EQ);
MOV(RA, Addr);
SetJumpTarget(DoNotWrite);
MOV(ValueReg, RS);
MOVI2R(Function, (u32)&Memory::Write_U8);
PUSH(4, R0, R1, R2, R3);
MOV(R0, ValueReg);
MOV(R1, Addr);
BL(Function);
POP(4, R0, R1, R2, R3);
gpr.Unlock(ValueReg, Addr, Function);
}
void JitArm::lwzx(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(LoadStore)
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
ARMReg RB = gpr.R(inst.RB);
ARMReg RD = gpr.R(inst.RD);
LDR(rA, R9, PPCSTATE_OFF(Exceptions));
CMP(rA, EXCEPTION_DSI);
FixupBranch DoNotLoad = B_CC(CC_EQ);
{
if (inst.RA)
{
ARMReg RA = gpr.R(inst.RA);
ADD(rB, RA, RB);
}
else
MOV(rB, RB);
MOVI2R(rA, (u32)&Memory::Read_U32);
PUSH(4, R0, R1, R2, R3);
MOV(R0, rB);
BL(rA);
MOV(rA, R0);
POP(4, R0, R1, R2, R3);
MOV(RD, rA);
gpr.Unlock(rA, rB);
}
SetJumpTarget(DoNotLoad);
//// u32 temp = Memory::Read_U32(_inst.RA ? (m_GPR[_inst.RA] + m_GPR[_inst.RB]) : m_GPR[_inst.RB]);
}
void JitArm::lha(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(LoadStore)
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
ARMReg RD = gpr.R(inst.RD);
LDR(rA, R9, PPCSTATE_OFF(Exceptions));
CMP(rA, EXCEPTION_DSI);
FixupBranch DoNotLoad = B_CC(CC_EQ);
if (inst.RA)
{
MOVI2R(rB, inst.SIMM_16);
ARMReg RA = gpr.R(inst.RA);
ADD(rB, rB, RA);
}
else
MOVI2R(rB, (u32)inst.SIMM_16);
MOVI2R(rA, (u32)&Memory::Read_U16);
PUSH(4, R0, R1, R2, R3);
MOV(R0, rB);
BL(rA);
MOV(rA, R0);
SXTH(rA, rA);
POP(4, R0, R1, R2, R3);
MOV(RD, rA);
gpr.Unlock(rA, rB);
SetJumpTarget(DoNotLoad);
}
void JitArm::ps_sel(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITPairedOff);
FALLBACK_IF(inst.Rc);
u32 a = inst.FA, b = inst.FB, c = inst.FC, d = inst.FD;
ARMReg vA0 = fpr.R0(a);
ARMReg vA1 = fpr.R1(a);
ARMReg vB0 = fpr.R0(b);
ARMReg vB1 = fpr.R1(b);
ARMReg vC0 = fpr.R0(c);
ARMReg vC1 = fpr.R1(c);
ARMReg vD0 = fpr.R0(d, false);
ARMReg vD1 = fpr.R1(d, false);
VCMP(vA0);
VMRS(_PC);
FixupBranch GT0 = B_CC(CC_GE);
VMOV(vD0, vB0);
FixupBranch EQ0 = B();
SetJumpTarget(GT0);
VMOV(vD0, vC0);
SetJumpTarget(EQ0);
VCMP(vA1);
VMRS(_PC);
FixupBranch GT1 = B_CC(CC_GE);
VMOV(vD1, vB1);
FixupBranch EQ1 = B();
SetJumpTarget(GT1);
VMOV(vD1, vC1);
SetJumpTarget(EQ1);
}
void JitArm::lfs(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(LoadStoreFloating)
Default(inst); return;
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
LDR(rA, R9, STRUCT_OFF(PowerPC::ppcState, Exceptions));
CMP(rA, EXCEPTION_DSI);
FixupBranch DoNotLoad = B_CC(CC_EQ);
if (inst.RA)
{
MOVI2R(rB, inst.SIMM_16);
ARMReg RA = gpr.R(inst.RA);
ADD(rB, rB, RA);
}
else
MOVI2R(rB, (u32)inst.SIMM_16);
MOVI2R(rA, (u32)&Memory::Read_U32);
PUSH(4, R0, R1, R2, R3);
MOV(R0, rB);
BL(rA);
MOV(rA, R0);
POP(4, R0, R1, R2, R3);
ARMReg v0 = fpr.R0(inst.FD, false);
ARMReg v1 = fpr.R1(inst.FD, false);
VMOV(v0, rA, false);
VMOV(v1, rA, false);
gpr.Unlock(rA, rB);
SetJumpTarget(DoNotLoad);
}
void JitArm::stX(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITLoadStoreOff);
u32 a = inst.RA, b = inst.RB, s = inst.RS;
s32 offset = inst.SIMM_16;
u32 accessSize = 0;
s32 regOffset = -1;
bool update = false;
bool fastmem = false;
switch (inst.OPCD)
{
case 45: // sthu
update = true;
case 44: // sth
accessSize = 16;
break;
case 31:
switch (inst.SUBOP10)
{
case 183: // stwux
update = true;
case 151: // stwx
fastmem = true;
accessSize = 32;
regOffset = b;
break;
case 247: // stbux
update = true;
case 215: // stbx
accessSize = 8;
regOffset = b;
break;
case 439: // sthux
update = true;
case 407: // sthx
accessSize = 16;
regOffset = b;
break;
}
break;
case 37: // stwu
update = true;
case 36: // stw
fastmem = true;
accessSize = 32;
break;
case 39: // stbu
update = true;
case 38: // stb
accessSize = 8;
break;
}
SafeStoreFromReg(fastmem, update ? a : (a ? a : -1), s, regOffset, accessSize, offset);
if (update)
{
ARMReg rA = gpr.GetReg();
ARMReg RB;
ARMReg RA = gpr.R(a);
if (regOffset != -1)
RB = gpr.R(regOffset);
// Check for DSI exception prior to writing back address
LDR(rA, R9, PPCSTATE_OFF(Exceptions));
TST(rA, EXCEPTION_DSI);
FixupBranch DoNotWrite = B_CC(CC_NEQ);
if (a)
{
if (regOffset == -1)
{
MOVI2R(rA, offset);
ADD(RA, RA, rA);
}
else
{
ADD(RA, RA, RB);
}
}
else
{
if (regOffset == -1)
MOVI2R(RA, (u32)offset);
else
MOV(RA, RB);
}
SetJumpTarget(DoNotWrite);
gpr.Unlock(rA);
}
}
void JitArm::fctiwx(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITFloatingPointOff)
u32 b = inst.FB;
u32 d = inst.FD;
ARMReg vB = fpr.R0(b);
ARMReg vD = fpr.R0(d);
ARMReg V0 = fpr.GetReg();
ARMReg V1 = fpr.GetReg();
ARMReg V2 = fpr.GetReg();
ARMReg rA = gpr.GetReg();
ARMReg fpscrReg = gpr.GetReg();
FixupBranch DoneMax, DoneMin;
LDR(fpscrReg, R9, PPCSTATE_OFF(fpscr));
MOVI2R(rA, (u32)minmaxFloat);
// Check if greater than max float
{
VLDR(V0, rA, 8); // Load Max
VCMPE(vB, V0);
VMRS(_PC); // Loads in to APSR
FixupBranch noException = B_CC(CC_LE);
VMOV(vD, V0); // Set to max
SetFPException(fpscrReg, FPSCR_VXCVI);
DoneMax = B();
SetJumpTarget(noException);
}
// Check if less than min float
{
VLDR(V0, rA, 0);
VCMPE(vB, V0);
VMRS(_PC);
FixupBranch noException = B_CC(CC_GE);
VMOV(vD, V0);
SetFPException(fpscrReg, FPSCR_VXCVI);
DoneMin = B();
SetJumpTarget(noException);
}
// Within ranges, convert to integer
// Set rounding mode first
// PPC <-> ARM rounding modes
// 0, 1, 2, 3 <-> 0, 3, 1, 2
ARMReg rB = gpr.GetReg();
VMRS(rA);
// Bits 22-23
BIC(rA, rA, Operand2(3, 5));
LDR(rB, R9, PPCSTATE_OFF(fpscr));
AND(rB, rB, 0x3); // Get the FPSCR rounding bits
CMP(rB, 1);
SetCC(CC_EQ); // zero
ORR(rA, rA, Operand2(3, 5));
SetCC(CC_NEQ);
CMP(rB, 2); // +inf
SetCC(CC_EQ);
ORR(rA, rA, Operand2(1, 5));
SetCC(CC_NEQ);
CMP(rB, 3); // -inf
SetCC(CC_EQ);
ORR(rA, rA, Operand2(2, 5));
SetCC();
VMSR(rA);
ORR(rA, rA, Operand2(3, 5));
VCVT(vD, vB, TO_INT | IS_SIGNED);
VMSR(rA);
gpr.Unlock(rB);
VCMPE(vD, vB);
VMRS(_PC);
SetCC(CC_EQ);
BIC(fpscrReg, fpscrReg, FRFIMask);
FixupBranch DoneEqual = B();
SetCC();
SetFPException(fpscrReg, FPSCR_XX);
ORR(fpscrReg, fpscrReg, FIMask);
VABS(V1, vB);
VABS(V2, vD);
VCMPE(V2, V1);
VMRS(_PC);
SetCC(CC_GT);
ORR(fpscrReg, fpscrReg, FRMask);
SetCC();
SetJumpTarget(DoneEqual);
SetJumpTarget(DoneMax);
SetJumpTarget(DoneMin);
MOVI2R(rA, (u32)&doublenum);
VLDR(V0, rA, 0);
NEONXEmitter nemit(this);
nemit.VORR(vD, vD, V0);
if (inst.Rc) Helper_UpdateCR1(fpscrReg, rA);
STR(fpscrReg, R9, PPCSTATE_OFF(fpscr));
gpr.Unlock(rA);
gpr.Unlock(fpscrReg);
fpr.Unlock(V0);
fpr.Unlock(V1);
fpr.Unlock(V2);
}
void JitArm::fctiwzx(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITFloatingPointOff)
u32 b = inst.FB;
u32 d = inst.FD;
ARMReg vB = fpr.R0(b);
ARMReg vD = fpr.R0(d);
ARMReg V0 = fpr.GetReg();
ARMReg V1 = fpr.GetReg();
ARMReg V2 = fpr.GetReg();
ARMReg rA = gpr.GetReg();
ARMReg fpscrReg = gpr.GetReg();
FixupBranch DoneMax, DoneMin;
LDR(fpscrReg, R9, PPCSTATE_OFF(fpscr));
MOVI2R(rA, (u32)minmaxFloat);
// Check if greater than max float
{
VLDR(V0, rA, 8); // Load Max
VCMPE(vB, V0);
VMRS(_PC); // Loads in to APSR
FixupBranch noException = B_CC(CC_LE);
VMOV(vD, V0); // Set to max
SetFPException(fpscrReg, FPSCR_VXCVI);
DoneMax = B();
SetJumpTarget(noException);
}
// Check if less than min float
{
VLDR(V0, rA, 0);
VCMPE(vB, V0);
VMRS(_PC);
FixupBranch noException = B_CC(CC_GE);
VMOV(vD, V0);
SetFPException(fpscrReg, FPSCR_VXCVI);
DoneMin = B();
SetJumpTarget(noException);
}
// Within ranges, convert to integer
VCVT(vD, vB, TO_INT | IS_SIGNED | ROUND_TO_ZERO);
VCMPE(vD, vB);
VMRS(_PC);
SetCC(CC_EQ);
BIC(fpscrReg, fpscrReg, FRFIMask);
FixupBranch DoneEqual = B();
SetCC();
SetFPException(fpscrReg, FPSCR_XX);
ORR(fpscrReg, fpscrReg, FIMask);
VABS(V1, vB);
VABS(V2, vD);
VCMPE(V2, V1);
VMRS(_PC);
SetCC(CC_GT);
ORR(fpscrReg, fpscrReg, FRMask);
SetCC();
SetJumpTarget(DoneEqual);
SetJumpTarget(DoneMax);
SetJumpTarget(DoneMin);
MOVI2R(rA, (u32)&doublenum);
VLDR(V0, rA, 0);
NEONXEmitter nemit(this);
nemit.VORR(vD, vD, V0);
if (inst.Rc) Helper_UpdateCR1(fpscrReg, rA);
STR(fpscrReg, R9, PPCSTATE_OFF(fpscr));
gpr.Unlock(rA);
gpr.Unlock(fpscrReg);
fpr.Unlock(V0);
fpr.Unlock(V1);
fpr.Unlock(V2);
}
void Jit::Comp_SVQ(MIPSOpcode op)
{
CONDITIONAL_DISABLE;
int imm = (signed short)(op&0xFFFC);
int vt = (((op >> 16) & 0x1f)) | ((op&1) << 5);
MIPSGPReg rs = _RS;
bool doCheck = false;
switch (op >> 26)
{
case 54: //lv.q
{
// CC might be set by slow path below, so load regs first.
u8 vregs[4];
GetVectorRegs(vregs, V_Quad, vt);
fpr.MapRegsAndSpillLockV(vregs, V_Quad, MAP_DIRTY | MAP_NOINIT);
if (gpr.IsImm(rs)) {
u32 addr = (imm + gpr.GetImm(rs)) & 0x3FFFFFFF;
MOVI2R(R0, addr + (u32)Memory::base);
} else {
gpr.MapReg(rs);
if (g_Config.bFastMemory) {
SetR0ToEffectiveAddress(rs, imm);
} else {
SetCCAndR0ForSafeAddress(rs, imm, R1);
doCheck = true;
}
ADD(R0, R0, R11);
}
#ifdef __ARM_ARCH_7S__
FixupBranch skip;
if (doCheck) {
skip = B_CC(CC_EQ);
}
for (int i = 0; i < 4; i++)
VLDR(fpr.V(vregs[i]), R0, i * 4);
if (doCheck) {
SetJumpTarget(skip);
SetCC(CC_AL);
}
#else
for (int i = 0; i < 4; i++)
VLDR(fpr.V(vregs[i]), R0, i * 4);
if (doCheck) {
SetCC(CC_EQ);
MOVI2R(R0, 0);
for (int i = 0; i < 4; i++)
VMOV(fpr.V(vregs[i]), R0);
SetCC(CC_AL);
}
#endif
}
break;
case 62: //sv.q
{
// CC might be set by slow path below, so load regs first.
u8 vregs[4];
GetVectorRegs(vregs, V_Quad, vt);
fpr.MapRegsAndSpillLockV(vregs, V_Quad, 0);
if (gpr.IsImm(rs)) {
u32 addr = (imm + gpr.GetImm(rs)) & 0x3FFFFFFF;
MOVI2R(R0, addr + (u32)Memory::base);
} else {
gpr.MapReg(rs);
if (g_Config.bFastMemory) {
SetR0ToEffectiveAddress(rs, imm);
} else {
SetCCAndR0ForSafeAddress(rs, imm, R1);
doCheck = true;
}
ADD(R0, R0, R11);
}
#ifdef __ARM_ARCH_7S__
FixupBranch skip;
if (doCheck) {
skip = B_CC(CC_EQ);
}
for (int i = 0; i < 4; i++)
VSTR(fpr.V(vregs[i]), R0, i * 4);
if (doCheck) {
SetJumpTarget(skip);
SetCC(CC_AL);
}
#else
for (int i = 0; i < 4; i++)
VSTR(fpr.V(vregs[i]), R0, i * 4);
if (doCheck) {
SetCC(CC_AL);
}
#endif
}
break;
default:
DISABLE;
break;
}
fpr.ReleaseSpillLocksAndDiscardTemps();
}
void Jit::Comp_SV(MIPSOpcode op) {
CONDITIONAL_DISABLE;
s32 imm = (signed short)(op&0xFFFC);
int vt = ((op >> 16) & 0x1f) | ((op & 3) << 5);
MIPSGPReg rs = _RS;
bool doCheck = false;
switch (op >> 26)
{
case 50: //lv.s // VI(vt) = Memory::Read_U32(addr);
{
// CC might be set by slow path below, so load regs first.
fpr.MapRegV(vt, MAP_DIRTY | MAP_NOINIT);
if (gpr.IsImm(rs)) {
u32 addr = (imm + gpr.GetImm(rs)) & 0x3FFFFFFF;
MOVI2R(R0, addr + (u32)Memory::base);
} else {
gpr.MapReg(rs);
if (g_Config.bFastMemory) {
SetR0ToEffectiveAddress(rs, imm);
} else {
SetCCAndR0ForSafeAddress(rs, imm, R1);
doCheck = true;
}
ADD(R0, R0, R11);
}
#ifdef __ARM_ARCH_7S__
FixupBranch skip;
if (doCheck) {
skip = B_CC(CC_EQ);
}
VLDR(fpr.V(vt), R0, 0);
if (doCheck) {
SetJumpTarget(skip);
SetCC(CC_AL);
}
#else
VLDR(fpr.V(vt), R0, 0);
if (doCheck) {
SetCC(CC_EQ);
MOVI2F(fpr.V(vt), 0.0f, R0);
SetCC(CC_AL);
}
#endif
}
break;
case 58: //sv.s // Memory::Write_U32(VI(vt), addr);
{
// CC might be set by slow path below, so load regs first.
fpr.MapRegV(vt);
if (gpr.IsImm(rs)) {
u32 addr = (imm + gpr.GetImm(rs)) & 0x3FFFFFFF;
MOVI2R(R0, addr + (u32)Memory::base);
} else {
gpr.MapReg(rs);
if (g_Config.bFastMemory) {
SetR0ToEffectiveAddress(rs, imm);
} else {
SetCCAndR0ForSafeAddress(rs, imm, R1);
doCheck = true;
}
ADD(R0, R0, R11);
}
#ifdef __ARM_ARCH_7S__
FixupBranch skip;
if (doCheck) {
skip = B_CC(CC_EQ);
}
VSTR(fpr.V(vt), R0, 0);
if (doCheck) {
SetJumpTarget(skip);
SetCC(CC_AL);
}
#else
VSTR(fpr.V(vt), R0, 0);
if (doCheck) {
SetCC(CC_AL);
}
#endif
}
break;
default:
DISABLE;
}
}
void JitArm::lfXX(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITLoadStoreFloatingOff);
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
ARMReg RA;
u32 a = inst.RA, b = inst.RB;
s32 offset = inst.SIMM_16;
bool single = false;
bool update = false;
bool zeroA = false;
s32 offsetReg = -1;
switch (inst.OPCD)
{
case 31:
switch (inst.SUBOP10)
{
case 567: // lfsux
single = true;
update = true;
offsetReg = b;
break;
case 535: // lfsx
single = true;
zeroA = true;
offsetReg = b;
break;
case 631: // lfdux
update = true;
offsetReg = b;
break;
case 599: // lfdx
zeroA = true;
offsetReg = b;
break;
}
break;
case 49: // lfsu
update = true;
single = true;
break;
case 48: // lfs
single = true;
zeroA = true;
break;
case 51: // lfdu
update = true;
break;
case 50: // lfd
zeroA = true;
break;
}
ARMReg v0 = fpr.R0(inst.FD), v1;
if (single)
v1 = fpr.R1(inst.FD);
if (update)
{
RA = gpr.R(a);
// Update path /always/ uses RA
if (offsetReg == -1) // uses SIMM_16
{
MOVI2R(rB, offset);
ADD(rB, rB, RA);
}
else
{
ADD(rB, gpr.R(offsetReg), RA);
}
}
else
{
if (zeroA)
{
if (offsetReg == -1)
{
if (a)
{
RA = gpr.R(a);
MOVI2R(rB, offset);
ADD(rB, rB, RA);
}
else
{
MOVI2R(rB, (u32)offset);
}
}
else
{
ARMReg RB = gpr.R(offsetReg);
if (a)
{
RA = gpr.R(a);
ADD(rB, RB, RA);
}
else
{
MOV(rB, RB);
}
}
}
}
LDR(rA, R9, PPCSTATE_OFF(Exceptions));
CMP(rA, EXCEPTION_DSI);
FixupBranch DoNotLoad = B_CC(CC_EQ);
if (update)
MOV(RA, rB);
if (Core::g_CoreStartupParameter.bFastmem)
{
Operand2 mask(2, 1); // ~(Memory::MEMVIEW32_MASK)
BIC(rB, rB, mask); // 1
MOVI2R(rA, (u32)Memory::base, false); // 2-3
ADD(rB, rB, rA); // 4
NEONXEmitter nemit(this);
if (single)
{
VLDR(S0, rB, 0);
nemit.VREV32(I_8, D0, D0); // Byte swap to result
VCVT(v0, S0, 0);
VCVT(v1, S0, 0);
}
else
{
VLDR(v0, rB, 0);
nemit.VREV64(I_8, v0, v0); // Byte swap to result
}
}
else
{
PUSH(4, R0, R1, R2, R3);
MOV(R0, rB);
if (single)
{
MOVI2R(rA, (u32)&Memory::Read_U32);
BL(rA);
VMOV(S0, R0);
VCVT(v0, S0, 0);
VCVT(v1, S0, 0);
}
else
{
MOVI2R(rA, (u32)&Memory::Read_F64);
BL(rA);
#if !defined(__ARM_PCS_VFP) // SoftFP returns in R0 and R1
VMOV(v0, R0);
#else
VMOV(v0, D0);
#endif
}
POP(4, R0, R1, R2, R3);
}
gpr.Unlock(rA, rB);
SetJumpTarget(DoNotLoad);
}
void Jit::Comp_FPULS(u32 op)
{
CONDITIONAL_DISABLE;
s32 offset = (s16)(op & 0xFFFF);
int ft = _FT;
int rs = _RS;
// u32 addr = R(rs) + offset;
// logBlocks = 1;
bool doCheck = false;
switch(op >> 26)
{
case 49: //FI(ft) = Memory::Read_U32(addr); break; //lwc1
fpr.SpillLock(ft);
fpr.MapReg(ft, MAP_NOINIT | MAP_DIRTY);
if (gpr.IsImm(rs)) {
u32 addr = (offset + gpr.GetImm(rs)) & 0x3FFFFFFF;
MOVI2R(R0, addr + (u32)Memory::base);
} else {
gpr.MapReg(rs);
if (g_Config.bFastMemory) {
SetR0ToEffectiveAddress(rs, offset);
} else {
SetCCAndR0ForSafeAddress(rs, offset, R1);
doCheck = true;
}
ADD(R0, R0, R11);
}
#ifdef __ARM_ARCH_7S__
FixupBranch skip;
if (doCheck) {
skip = B_CC(CC_EQ);
}
VLDR(fpr.R(ft), R0, 0);
if (doCheck) {
SetJumpTarget(skip);
SetCC(CC_AL);
}
#else
VLDR(fpr.R(ft), R0, 0);
if (doCheck) {
SetCC(CC_EQ);
MOVI2R(R0, 0);
VMOV(fpr.R(ft), R0);
SetCC(CC_AL);
}
#endif
fpr.ReleaseSpillLock(ft);
break;
case 57: //Memory::Write_U32(FI(ft), addr); break; //swc1
fpr.MapReg(ft);
if (gpr.IsImm(rs)) {
u32 addr = (offset + gpr.GetImm(rs)) & 0x3FFFFFFF;
MOVI2R(R0, addr + (u32)Memory::base);
} else {
gpr.MapReg(rs);
if (g_Config.bFastMemory) {
SetR0ToEffectiveAddress(rs, offset);
} else {
SetCCAndR0ForSafeAddress(rs, offset, R1);
doCheck = true;
}
ADD(R0, R0, R11);
}
#ifdef __ARM_ARCH_7S__
FixupBranch skip2;
if (doCheck) {
skip2 = B_CC(CC_EQ);
}
VSTR(fpr.R(ft), R0, 0);
if (doCheck) {
SetJumpTarget(skip2);
SetCC(CC_AL);
}
#else
VSTR(fpr.R(ft), R0, 0);
if (doCheck) {
SetCC(CC_AL);
}
#endif
break;
default:
Comp_Generic(op);
return;
}
}
void JitArm::lXX(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITLoadStoreOff);
u32 a = inst.RA, b = inst.RB, d = inst.RD;
s32 offset = inst.SIMM_16;
u32 accessSize = 0;
s32 offsetReg = -1;
bool update = false;
bool signExtend = false;
bool reverse = false;
bool fastmem = false;
switch (inst.OPCD)
{
case 31:
switch (inst.SUBOP10)
{
case 55: // lwzux
update = true;
case 23: // lwzx
fastmem = true;
accessSize = 32;
offsetReg = b;
break;
case 119: //lbzux
update = true;
case 87: // lbzx
fastmem = true;
accessSize = 8;
offsetReg = b;
break;
case 311: // lhzux
update = true;
case 279: // lhzx
fastmem = true;
accessSize = 16;
offsetReg = b;
break;
case 375: // lhaux
update = true;
case 343: // lhax
accessSize = 16;
signExtend = true;
offsetReg = b;
break;
case 534: // lwbrx
accessSize = 32;
reverse = true;
break;
case 790: // lhbrx
accessSize = 16;
reverse = true;
break;
}
break;
case 33: // lwzu
update = true;
case 32: // lwz
fastmem = true;
accessSize = 32;
break;
case 35: // lbzu
update = true;
case 34: // lbz
fastmem = true;
accessSize = 8;
break;
case 41: // lhzu
update = true;
case 40: // lhz
fastmem = true;
accessSize = 16;
break;
case 43: // lhau
update = true;
case 42: // lha
signExtend = true;
accessSize = 16;
break;
}
// Check for exception before loading
ARMReg rA = gpr.GetReg(false);
LDR(rA, R9, PPCSTATE_OFF(Exceptions));
TST(rA, EXCEPTION_DSI);
FixupBranch DoNotLoad = B_CC(CC_NEQ);
SafeLoadToReg(fastmem, d, update ? a : (a ? a : -1), offsetReg, accessSize, offset, signExtend, reverse);
if (update)
{
ARMReg RA = gpr.R(a);
if (offsetReg == -1)
{
rA = gpr.GetReg(false);
MOVI2R(rA, offset);
ADD(RA, RA, rA);
}
else
{
ADD(RA, RA, gpr.R(offsetReg));
}
}
SetJumpTarget(DoNotLoad);
// LWZ idle skipping
if (SConfig::GetInstance().m_LocalCoreStartupParameter.bSkipIdle &&
inst.OPCD == 32 &&
(inst.hex & 0xFFFF0000) == 0x800D0000 &&
(Memory::ReadUnchecked_U32(js.compilerPC + 4) == 0x28000000 ||
(SConfig::GetInstance().m_LocalCoreStartupParameter.bWii && Memory::ReadUnchecked_U32(js.compilerPC + 4) == 0x2C000000)) &&
Memory::ReadUnchecked_U32(js.compilerPC + 8) == 0x4182fff8)
{
ARMReg RD = gpr.R(d);
// if it's still 0, we can wait until the next event
TST(RD, RD);
FixupBranch noIdle = B_CC(CC_NEQ);
gpr.Flush(FLUSH_MAINTAIN_STATE);
fpr.Flush(FLUSH_MAINTAIN_STATE);
rA = gpr.GetReg();
MOVI2R(rA, (u32)&PowerPC::OnIdle);
MOVI2R(R0, PowerPC::ppcState.gpr[a] + (s32)(s16)inst.SIMM_16);
BL(rA);
gpr.Unlock(rA);
WriteExceptionExit();
SetJumpTarget(noIdle);
//js.compilerPC += 8;
return;
}
}
void JitArmILAsmRoutineManager::Generate()
{
enterCode = GetCodePtr();
PUSH(9, R4, R5, R6, R7, R8, R9, R10, R11, _LR);
// Take care to 8-byte align stack for function calls.
// We are misaligned here because of an odd number of args for PUSH.
// It's not like x86 where you need to account for an extra 4 bytes
// consumed by CALL.
SUB(_SP, _SP, 4);
MOVI2R(R0, (u32)&CoreTiming::downcount);
MOVI2R(R9, (u32)&PowerPC::ppcState.spr[0]);
FixupBranch skipToRealDispatcher = B();
dispatcher = GetCodePtr();
printf("ILDispatcher is %p\n", dispatcher);
// Downcount Check
// The result of slice decrementation should be in flags if somebody jumped here
// IMPORTANT - We jump on negative, not carry!!!
FixupBranch bail = B_CC(CC_MI);
SetJumpTarget(skipToRealDispatcher);
dispatcherNoCheck = GetCodePtr();
// This block of code gets the address of the compiled block of code
// It runs though to the compiling portion if it isn't found
LDR(R12, R9, PPCSTATE_OFF(pc));// Load the current PC into R12
Operand2 iCacheMask = Operand2(0xE, 2); // JIT_ICACHE_MASK
BIC(R12, R12, iCacheMask); // R12 contains PC & JIT_ICACHE_MASK here.
MOVI2R(R14, (u32)jit->GetBlockCache()->iCache);
LDR(R12, R14, R12); // R12 contains iCache[PC & JIT_ICACHE_MASK] here
// R12 Confirmed this is the correct iCache Location loaded.
TST(R12, 0x80); // Test to see if it is a JIT block.
SetCC(CC_EQ);
// Success, it is our Jitblock.
MOVI2R(R14, (u32)jit->GetBlockCache()->GetCodePointers());
// LDR R14 right here to get CodePointers()[0] pointer.
LSL(R12, R12, 2); // Multiply by four because address locations are u32 in size
LDR(R14, R14, R12); // Load the block address in to R14
B(R14);
// No need to jump anywhere after here, the block will go back to dispatcher start
SetCC();
// If we get to this point, that means that we don't have the block cached to execute
// So call ArmJit to compile the block and then execute it.
MOVI2R(R14, (u32)&Jit);
BL(R14);
B(dispatcherNoCheck);
// fpException()
// Floating Point Exception Check, Jumped to if false
fpException = GetCodePtr();
LDR(R0, R9, PPCSTATE_OFF(Exceptions));
ORR(R0, R0, EXCEPTION_FPU_UNAVAILABLE);
STR(R0, R9, PPCSTATE_OFF(Exceptions));
QuickCallFunction(R14, (void*)&PowerPC::CheckExceptions);
LDR(R0, R9, PPCSTATE_OFF(npc));
STR(R0, R9, PPCSTATE_OFF(pc));
B(dispatcher);
SetJumpTarget(bail);
doTiming = GetCodePtr();
// XXX: In JIT64, Advance() gets called /after/ the exception checking
// once it jumps back to the start of outerLoop
QuickCallFunction(R14, (void*)&CoreTiming::Advance);
// Does exception checking
testExceptions = GetCodePtr();
LDR(R0, R9, PPCSTATE_OFF(pc));
STR(R0, R9, PPCSTATE_OFF(npc));
QuickCallFunction(R14, (void*)&PowerPC::CheckExceptions);
LDR(R0, R9, PPCSTATE_OFF(npc));
STR(R0, R9, PPCSTATE_OFF(pc));
// Check the state pointer to see if we are exiting
// Gets checked on every exception check
MOVI2R(R0, (u32)PowerPC::GetStatePtr());
MVN(R1, 0);
LDR(R0, R0);
TST(R0, R1);
FixupBranch Exit = B_CC(CC_NEQ);
B(dispatcher);
SetJumpTarget(Exit);
ADD(_SP, _SP, 4);
POP(9, R4, R5, R6, R7, R8, R9, R10, R11, _PC); // Returns
GenerateCommon();
FlushIcache();
}
