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::ps_muls1(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITPairedOff);
FALLBACK_IF(inst.Rc);
u32 a = inst.FA, c = inst.FC, d = inst.FD;
ARMReg vA0 = fpr.R0(a);
ARMReg vA1 = fpr.R1(a);
ARMReg vC1 = fpr.R1(c);
ARMReg vD0 = fpr.R0(d, false);
ARMReg vD1 = fpr.R1(d, false);
ARMReg V0 = fpr.GetReg();
ARMReg V1 = fpr.GetReg();
VMUL(V0, vA0, vC1);
VMUL(V1, vA1, vC1);
VMOV(vD0, V0);
VMOV(vD1, V1);
fpr.Unlock(V0);
fpr.Unlock(V1);
}
void JitArm::ps_mr(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);
VMOV(vD0, vB0);
VMOV(vD1, vB1);
}
void JitArm::stfs(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITLoadStoreFloatingOff);
ARMReg rA = gpr.GetReg();
ARMReg rB = gpr.GetReg();
ARMReg v0 = fpr.R0(inst.FS);
VCVT(S0, v0, 0);
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::Write_U32);
PUSH(4, R0, R1, R2, R3);
VMOV(R0, S0);
MOV(R1, rB);
BL(rA);
POP(4, R0, R1, R2, R3);
gpr.Unlock(rA, rB);
}
void JitArm::ps_merge10(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITPairedOff);
FALLBACK_IF(inst.Rc);
u32 a = inst.FA, b = inst.FB, d = inst.FD;
ARMReg vA1 = fpr.R1(a);
ARMReg vB0 = fpr.R0(b);
ARMReg vD0 = fpr.R0(d, false);
ARMReg vD1 = fpr.R1(d, false);
ARMReg V0 = fpr.GetReg();
VMOV(V0, vB0);
VMOV(vD0, vA1);
VMOV(vD1, V0);
fpr.Unlock(V0);
}
void JitArm::fmrx(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(FloatingPoint)
ARMReg vD = fpr.R0(inst.FD);
ARMReg vB = fpr.R0(inst.FB);
VMOV(vD, vB);
if (inst.Rc) Helper_UpdateCR1(vD);
}
void Jit::ApplyPrefixST(u8 *vregs, u32 prefix, VectorSize sz) {
if (prefix == 0xE4) return;
int n = GetNumVectorElements(sz);
u8 origV[4];
static const float constantArray[8] = {0.f, 1.f, 2.f, 0.5f, 3.f, 1.f/3.f, 0.25f, 1.f/6.f};
for (int i = 0; i < n; i++)
origV[i] = vregs[i];
for (int i = 0; i < n; i++)
{
int regnum = (prefix >> (i*2)) & 3;
int abs = (prefix >> (8+i)) & 1;
int negate = (prefix >> (16+i)) & 1;
int constants = (prefix >> (12+i)) & 1;
// Unchanged, hurray.
if (!constants && regnum == i && !abs && !negate)
continue;
// This puts the value into a temp reg, so we won't write the modified value back.
vregs[i] = fpr.GetTempV();
fpr.MapRegV(vregs[i], MAP_NOINIT | MAP_DIRTY);
if (!constants) {
// Prefix may say "z, z, z, z" but if this is a pair, we force to x.
// TODO: But some ops seem to use const 0 instead?
if (regnum >= n) {
ERROR_LOG_REPORT(CPU, "Invalid VFPU swizzle: %08x / %d", prefix, sz);
regnum = 0;
}
if (abs) {
VABS(fpr.V(vregs[i]), fpr.V(origV[regnum]));
} else {
VMOV(fpr.V(vregs[i]), fpr.V(origV[regnum]));
}
} else {
// TODO: There is VMOV s, imm on ARM, that can generate some of these constants. Not 1/3 or 1/6 though.
MOVI2F(fpr.V(vregs[i]), constantArray[regnum + (abs<<2)], R0);
}
// TODO: This can be integrated into the VABS / VMOV above, and also the constants.
if (negate)
VNEG(fpr.V(vregs[i]), fpr.V(vregs[i]));
// TODO: This probably means it will swap out soon, inefficiently...
fpr.ReleaseSpillLockV(vregs[i]);
}
}
void Jit::ApplyPrefixST(u8 *vregs, u32 prefix, VectorSize sz) {
if (prefix == 0xE4) return;
int n = GetNumVectorElements(sz);
u8 origV[4];
static const float constantArray[8] = {0.f, 1.f, 2.f, 0.5f, 3.f, 1.f/3.f, 0.25f, 1.f/6.f};
for (int i = 0; i < n; i++)
origV[i] = vregs[i];
for (int i = 0; i < n; i++)
{
int regnum = (prefix >> (i*2)) & 3;
int abs = (prefix >> (8+i)) & 1;
int negate = (prefix >> (16+i)) & 1;
int constants = (prefix >> (12+i)) & 1;
// Unchanged, hurray.
if (!constants && regnum == i && !abs && !negate)
continue;
// This puts the value into a temp reg, so we won't write the modified value back.
vregs[i] = fpr.GetTempV();
if (!constants) {
fpr.MapDirtyInV(vregs[i], origV[regnum]);
fpr.SpillLockV(vregs[i]);
// Prefix may say "z, z, z, z" but if this is a pair, we force to x.
// TODO: But some ops seem to use const 0 instead?
if (regnum >= n) {
WARN_LOG(CPU, "JIT: Invalid VFPU swizzle: %08x : %d / %d at PC = %08x (%s)", prefix, regnum, n, js.compilerPC, currentMIPS->DisasmAt(js.compilerPC));
regnum = 0;
}
if (abs) {
VABS(fpr.V(vregs[i]), fpr.V(origV[regnum]));
if (negate)
VNEG(fpr.V(vregs[i]), fpr.V(vregs[i]));
} else {
if (negate)
VNEG(fpr.V(vregs[i]), fpr.V(origV[regnum]));
else
VMOV(fpr.V(vregs[i]), fpr.V(origV[regnum]));
}
} else {
fpr.MapRegV(vregs[i], MAP_DIRTY | MAP_NOINIT);
fpr.SpillLockV(vregs[i]);
MOVI2F(fpr.V(vregs[i]), constantArray[regnum + (abs<<2)], R0, negate);
}
}
}
void JitArm::fsubsx(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(FloatingPoint)
ARMReg vA = fpr.R0(inst.FA);
ARMReg vB = fpr.R0(inst.FB);
ARMReg vD0 = fpr.R0(inst.FD);
ARMReg vD1 = fpr.R1(inst.FD);
VSUB(vD0, vA, vB);
VMOV(vD1, vD0);
if (inst.Rc) Helper_UpdateCR1(vD0);
}
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::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);
}
// Breaks Animal Crossing
void JitArm::fmulsx(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(FloatingPoint)
Default(inst); return;
ARMReg vA = fpr.R0(inst.FA);
ARMReg vC = fpr.R0(inst.FC);
ARMReg vD0 = fpr.R0(inst.FD);
ARMReg vD1 = fpr.R1(inst.FD);
VMUL(vD0, vA, vC);
VMOV(vD1, vD0);
if (inst.Rc) Helper_UpdateCR1(vD0);
}
void JitArm::ps_sum1(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 vB1 = fpr.R1(b);
ARMReg vC0 = fpr.R0(c);
ARMReg vD0 = fpr.R0(d, false);
ARMReg vD1 = fpr.R1(d, false);
VMOV(vD0, vC0);
VADD(vD1, vA0, vB1);
}
void Jit::Comp_VDot(u32 op)
{
// DISABLE;
CONDITIONAL_DISABLE;
// WARNING: No prefix support!
if (js.MayHavePrefix()) {
Comp_Generic(op);
js.EatPrefix();
return;
}
int vd = _VD;
int vs = _VS;
int vt = _VT;
VectorSize sz = GetVecSize(op);
// TODO: Force read one of them into regs? probably not.
u8 sregs[4], tregs[4];
GetVectorRegs(sregs, sz, vs);
GetVectorRegs(tregs, sz, vt);
// TODO: applyprefixST here somehow (shuffle, etc...)
fpr.MapRegsV(sregs, sz, 0);
fpr.MapRegsV(tregs, sz, 0);
VMUL(S0, fpr.V(sregs[0]), fpr.V(tregs[0]));
int n = GetNumVectorElements(sz);
for (int i = 1; i < n; i++) {
// sum += s[i]*t[i];
VMLA(S0, fpr.V(sregs[i]), fpr.V(tregs[i]));
}
fpr.ReleaseSpillLocks();
fpr.MapRegV(vd, MAP_NOINIT | MAP_DIRTY);
// TODO: applyprefixD here somehow (write mask etc..)
VMOV(fpr.V(vd), S0);
fpr.ReleaseSpillLocks();
js.EatPrefix();
}
void kamikaze_add(vec3f pos, float s)
{
actor* a = ACTOR_get(actor_pool);
VMOV(a->pos, pos);
a->pos[2] -= 0.1f;
a->vel[0] = 0.0f;
a->vel[1] = -rand01();
a->vel[2] = -5.0f -10.0f*rand01();
a->ang = 0.0;
a->collide_size[0] = s;
a->collide_size[1] = s;
a->aux[0] = linear(20.0f, 30.0f,rand01());
a->update = kamikaze_update;
a->render = kamikaze_render;
}
void Jit::Comp_VVectorInit(u32 op)
{
CONDITIONAL_DISABLE;
// WARNING: No prefix support!
if (js.MayHavePrefix()) {
Comp_Generic(op);
js.EatPrefix();
return;
}
switch ((op >> 16) & 0xF)
{
case 6: // v=zeros; break; //vzero
MOVI2F(S0, 0.0f, R0);
break;
case 7: // v=ones; break; //vone
MOVI2F(S0, 1.0f, R0);
break;
default:
DISABLE;
break;
}
VectorSize sz = GetVecSize(op);
int n = GetNumVectorElements(sz);
u8 dregs[4];
GetVectorRegsPrefixD(dregs, sz, _VD);
fpr.MapRegsV(dregs, sz, MAP_NOINIT | MAP_DIRTY);
for (int i = 0; i < n; ++i)
VMOV(fpr.V(dregs[i]), S0);
ApplyPrefixD(dregs, sz);
fpr.ReleaseSpillLocks();
}
void Jit::Comp_VVectorInit(MIPSOpcode op)
{
CONDITIONAL_DISABLE;
// WARNING: No prefix support!
if (js.HasUnknownPrefix() || disablePrefixes) {
DISABLE;
}
switch ((op >> 16) & 0xF)
{
case 6: // v=zeros; break; //vzero
MOVI2F(S0, 0.0f, R0);
break;
case 7: // v=ones; break; //vone
MOVI2F(S0, 1.0f, R0);
break;
default:
DISABLE;
break;
}
VectorSize sz = GetVecSize(op);
int n = GetNumVectorElements(sz);
u8 dregs[4];
GetVectorRegsPrefixD(dregs, sz, _VD);
fpr.MapRegsAndSpillLockV(dregs, sz, MAP_NOINIT | MAP_DIRTY);
for (int i = 0; i < n; ++i)
VMOV(fpr.V(dregs[i]), S0);
ApplyPrefixD(dregs, sz);
fpr.ReleaseSpillLocksAndDiscardTemps();
}
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.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);
}
VLDR(fpr.R(ft), R0, 0);
if (doCheck) {
SetCC(CC_EQ);
MOVI2R(R0, 0);
VMOV(fpr.R(ft), R0);
SetCC(CC_AL);
}
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);
}
VSTR(fpr.R(ft), R0, 0);
if (doCheck) {
SetCC(CC_AL);
}
break;
default:
Comp_Generic(op);
return;
}
}
void Jit::Comp_FPU2op(u32 op)
{
CONDITIONAL_DISABLE;
int fs = _FS;
int fd = _FD;
// logBlocks = 1;
switch (op & 0x3f)
{
case 4: //F(fd) = sqrtf(F(fs)); break; //sqrt
fpr.MapDirtyIn(fd, fs);
VSQRT(fpr.R(fd), fpr.R(fs));
break;
case 5: //F(fd) = fabsf(F(fs)); break; //abs
fpr.MapDirtyIn(fd, fs);
VABS(fpr.R(fd), fpr.R(fs));
break;
case 6: //F(fd) = F(fs); break; //mov
fpr.MapDirtyIn(fd, fs);
VMOV(fpr.R(fd), fpr.R(fs));
break;
case 7: //F(fd) = -F(fs); break; //neg
fpr.MapDirtyIn(fd, fs);
VNEG(fpr.R(fd), fpr.R(fs));
break;
case 12: //FsI(fd) = (int)floorf(F(fs)+0.5f); break; //round.w.s
fpr.MapDirtyIn(fd, fs);
VCVT(fpr.R(fd), fpr.R(fs), TO_INT | IS_SIGNED);
break;
case 13: //FsI(fd) = Rto0(F(fs))); break; //trunc.w.s
fpr.MapDirtyIn(fd, fs);
VCVT(fpr.R(fd), fpr.R(fs), TO_INT | IS_SIGNED | ROUND_TO_ZERO);
break;
case 14: //FsI(fd) = (int)ceilf (F(fs)); break; //ceil.w.s
fpr.MapDirtyIn(fd, fs);
MOVI2F(S0, 0.5f, R0);
VADD(S0,fpr.R(fs),S0);
VCVT(fpr.R(fd), S0, TO_INT | IS_SIGNED);
break;
case 15: //FsI(fd) = (int)floorf(F(fs)); break; //floor.w.s
fpr.MapDirtyIn(fd, fs);
MOVI2F(S0, 0.5f, R0);
VSUB(S0,fpr.R(fs),S0);
VCVT(fpr.R(fd), S0, TO_INT | IS_SIGNED);
break;
case 32: //F(fd) = (float)FsI(fs); break; //cvt.s.w
fpr.MapDirtyIn(fd, fs);
VCVT(fpr.R(fd), fpr.R(fs), TO_FLOAT | IS_SIGNED);
break;
case 36: //FsI(fd) = (int) F(fs); break; //cvt.w.s
fpr.MapDirtyIn(fd, fs);
LDR(R0, CTXREG, offsetof(MIPSState, fcr31));
AND(R0, R0, Operand2(3));
// MIPS Rounding Mode:
// 0: Round nearest
// 1: Round to zero
// 2: Round up (ceil)
// 3: Round down (floor)
CMP(R0, Operand2(2));
SetCC(CC_GE); MOVI2F(S0, 0.5f, R1);
SetCC(CC_GT); VSUB(S0,fpr.R(fs),S0);
SetCC(CC_EQ); VADD(S0,fpr.R(fs),S0);
SetCC(CC_GE); VCVT(fpr.R(fd), S0, TO_INT | IS_SIGNED); /* 2,3 */
SetCC(CC_AL);
CMP(R0, Operand2(1));
SetCC(CC_EQ); VCVT(fpr.R(fd), fpr.R(fs), TO_INT | IS_SIGNED | ROUND_TO_ZERO); /* 1 */
SetCC(CC_LT); VCVT(fpr.R(fd), fpr.R(fs), TO_INT | IS_SIGNED); /* 0 */
SetCC(CC_AL);
break;
default:
DISABLE;
}
}
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(u32 op) {
CONDITIONAL_DISABLE;
s32 imm = (signed short)(op&0xFFFC);
int vt = ((op >> 16) & 0x1f) | ((op & 3) << 5);
int 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);
fpr.ReleaseSpillLocks();
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);
}
VLDR(fpr.V(vt), R0, 0);
if (doCheck) {
SetCC(CC_EQ);
MOVI2R(R0, 0);
VMOV(fpr.V(vt), R0);
SetCC(CC_AL);
}
}
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);
fpr.ReleaseSpillLocks();
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);
}
VSTR(fpr.V(vt), R0, 0);
if (doCheck) {
SetCC(CC_AL);
}
}
break;
default:
DISABLE;
}
}
void Jit::ApplyPrefixD(const u8 *vregs, VectorSize sz) {
_assert_(js.prefixDFlag & ArmJitState::PREFIX_KNOWN);
if (!js.prefixD) return;
int n = GetNumVectorElements(sz);
for (int i = 0; i < n; i++) {
if (js.VfpuWriteMask(i))
continue;
// TODO: These clampers are wrong - put this into google
// and look at the plot: abs(x) - abs(x-0.5) + 0.5
// It's too steep.
// Also, they mishandle NaN and Inf.
int sat = (js.prefixD >> (i * 2)) & 3;
if (sat == 1) {
// clamped = fabs(x) - fabs(x-0.5f) + 0.5f; // [ 0, 1]
fpr.MapRegV(vregs[i], MAP_DIRTY);
MOVI2F(S0, 0.0f, R0);
MOVI2F(S1, 1.0f, R0);
VCMP(fpr.V(vregs[i]), S0);
VMRS_APSR(); // Move FP flags from FPSCR to APSR (regular flags).
SetCC(CC_LE);
VMOV(fpr.V(vregs[i]), S0);
SetCC(CC_AL);
VCMP(fpr.V(vregs[i]), S1);
VMRS_APSR(); // Move FP flags from FPSCR to APSR (regular flags).
SetCC(CC_GT);
VMOV(fpr.V(vregs[i]), S1);
SetCC(CC_AL);
/*
VABS(S1, fpr.V(vregs[i])); // S1 = fabs(x)
VSUB(fpr.V(vregs[i]), fpr.V(vregs[i]), S0); // S2 = fabs(x-0.5f) {VABD}
VABS(fpr.V(vregs[i]), fpr.V(vregs[i]));
VSUB(fpr.V(vregs[i]), S1, fpr.V(vregs[i])); // v[i] = S1 - S2 + 0.5f
VADD(fpr.V(vregs[i]), fpr.V(vregs[i]), S0);*/
} else if (sat == 3) {
fpr.MapRegV(vregs[i], MAP_DIRTY);
MOVI2F(S0, -1.0f, R0);
MOVI2F(S1, 1.0f, R0);
VCMP(fpr.V(vregs[i]), S0);
VMRS_APSR(); // Move FP flags from FPSCR to APSR (regular flags).
SetCC(CC_LT);
VMOV(fpr.V(vregs[i]), S0);
SetCC(CC_AL);
VCMP(fpr.V(vregs[i]), S1);
VMRS_APSR(); // Move FP flags from FPSCR to APSR (regular flags).
SetCC(CC_GT);
VMOV(fpr.V(vregs[i]), S1);
SetCC(CC_AL);
// clamped = fabs(x) - fabs(x-1.0f); // [-1, 1]
/*
fpr.MapRegV(vregs[i], MAP_DIRTY);
MOVI2F(S0, 1.0f, R0);
VABS(S1, fpr.V(vregs[i])); // S1 = fabs(x)
VSUB(fpr.V(vregs[i]), fpr.V(vregs[i]), S0); // S2 = fabs(x-1.0f) {VABD}
VABS(fpr.V(vregs[i]), fpr.V(vregs[i]));
VSUB(fpr.V(vregs[i]), S1, fpr.V(vregs[i])); // v[i] = S1 - S2
*/
}
}
}
void Jit::Comp_VV2Op(u32 op) {
CONDITIONAL_DISABLE;
DISABLE;
if (js.HasUnknownPrefix())
DISABLE;
VectorSize sz = GetVecSize(op);
int n = GetNumVectorElements(sz);
u8 sregs[4], dregs[4];
GetVectorRegsPrefixS(sregs, sz, _VS);
GetVectorRegsPrefixD(dregs, sz, _VD);
ARMReg tempxregs[4];
for (int i = 0; i < n; ++i)
{
if (!IsOverlapSafeAllowS(dregs[i], i, n, sregs))
{
int reg = fpr.GetTempV();
fpr.MapRegV(reg, MAP_NOINIT | MAP_DIRTY);
fpr.SpillLockV(reg);
tempxregs[i] = fpr.V(reg);
}
else
{
fpr.MapRegV(dregs[i], (dregs[i] == sregs[i] ? 0 : MAP_NOINIT) | MAP_DIRTY);
fpr.SpillLockV(dregs[i]);
tempxregs[i] = fpr.V(dregs[i]);
}
}
// Warning: sregs[i] and tempxregs[i] may be the same reg.
// Helps for vmov, hurts for vrcp, etc.
for (int i = 0; i < n; ++i)
{
switch ((op >> 16) & 0x1f)
{
case 0: // d[i] = s[i]; break; //vmov
// Probably for swizzle.
VMOV(tempxregs[i], fpr.V(sregs[i]));
break;
case 1: // d[i] = fabsf(s[i]); break; //vabs
//if (!fpr.V(sregs[i]).IsSimpleReg(tempxregs[i]))
VABS(tempxregs[i], fpr.V(sregs[i]));
break;
case 2: // d[i] = -s[i]; break; //vneg
VNEG(tempxregs[i], fpr.V(sregs[i]));
break;
case 4: // if (s[i] < 0) d[i] = 0; else {if(s[i] > 1.0f) d[i] = 1.0f; else d[i] = s[i];} break; // vsat0
DISABLE;
break;
case 5: // if (s[i] < -1.0f) d[i] = -1.0f; else {if(s[i] > 1.0f) d[i] = 1.0f; else d[i] = s[i];} break; // vsat1
DISABLE;
break;
case 16: // d[i] = 1.0f / s[i]; break; //vrcp
MOVI2F(S0, 1.0f, R0);
VDIV(tempxregs[i], S0, fpr.V(sregs[i]));
break;
case 17: // d[i] = 1.0f / sqrtf(s[i]); break; //vrsq
MOVI2F(S0, 1.0f, R0);
VSQRT(S1, fpr.V(sregs[i]));
VDIV(tempxregs[i], S0, S1);
break;
case 18: // d[i] = sinf((float)M_PI_2 * s[i]); break; //vsin
DISABLE;
break;
case 19: // d[i] = cosf((float)M_PI_2 * s[i]); break; //vcos
DISABLE;
break;
case 20: // d[i] = powf(2.0f, s[i]); break; //vexp2
DISABLE;
break;
case 21: // d[i] = logf(s[i])/log(2.0f); break; //vlog2
DISABLE;
break;
case 22: // d[i] = sqrtf(s[i]); break; //vsqrt
VSQRT(tempxregs[i], fpr.V(sregs[i]));
VABS(tempxregs[i], tempxregs[i]);
break;
case 23: // d[i] = asinf(s[i] * (float)M_2_PI); break; //vasin
DISABLE;
break;
case 24: // d[i] = -1.0f / s[i]; break; // vnrcp
MOVI2F(S0, -1.0f, R0);
VDIV(tempxregs[i], S0, fpr.V(sregs[i]));
break;
case 26: // d[i] = -sinf((float)M_PI_2 * s[i]); break; // vnsin
DISABLE;
break;
case 28: // d[i] = 1.0f / expf(s[i] * (float)M_LOG2E); break; // vrexp2
DISABLE;
break;
}
}
fpr.MapRegsV(dregs, sz, MAP_NOINIT | MAP_DIRTY);
for (int i = 0; i < n; ++i)
{
VMOV(fpr.V(dregs[i]), tempxregs[i]);
}
ApplyPrefixD(dregs, sz);
fpr.ReleaseSpillLocks();
}
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 JitArm::stfXX(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 663: // stfsx
single = true;
zeroA = true;
offsetReg = b;
break;
case 695: // stfsux
single = true;
offsetReg = b;
break;
case 727: // stfdx
zeroA = true;
offsetReg = b;
break;
case 759: // stfdux
update = true;
offsetReg = b;
break;
}
break;
case 53: // stfsu
update = true;
single = true;
break;
case 52: // stfs
single = true;
zeroA = true;
break;
case 55: // stfdu
update = true;
break;
case 54: // stfd
zeroA = true;
break;
}
ARMReg v0 = fpr.R0(inst.FS);
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);
}
}
}
}
if (update)
{
LDR(rA, R9, PPCSTATE_OFF(Exceptions));
CMP(rA, EXCEPTION_DSI);
SetCC(CC_NEQ);
MOV(RA, rB);
SetCC();
}
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)
{
VCVT(S0, v0, 0);
nemit.VREV32(I_8, D0, D0);
VSTR(S0, rB, 0);
}
else
{
nemit.VREV64(I_8, D0, v0);
VSTR(D0, rB, 0);
}
}
else
{
PUSH(4, R0, R1, R2, R3);
if (single)
{
MOVI2R(rA, (u32)&Memory::Write_U32);
VCVT(S0, v0, 0);
VMOV(R0, S0);
MOV(R1, rB);
BL(rA);
}
else
{
MOVI2R(rA, (u32)&Memory::Write_F64);
#if !defined(__ARM_PCS_VFP) // SoftFP returns in R0 and R1
VMOV(R0, v0);
MOV(R2, rB);
#else
VMOV(D0, v0);
MOV(R0, rB);
#endif
BL(rA);
}
POP(4, R0, R1, R2, R3);
}
gpr.Unlock(rA, rB);
}
void ArmJit::NEONApplyPrefixD(DestARMReg dest) {
// Apply clamps to dest.rd
int n = GetNumVectorElements(dest.sz);
int sat1_mask = 0;
int sat3_mask = 0;
int full_mask = (1 << n) - 1;
for (int i = 0; i < n; i++) {
int sat = (js.prefixD >> (i * 2)) & 3;
if (sat == 1)
sat1_mask |= 1 << i;
if (sat == 3)
sat3_mask |= 1 << i;
}
if (sat1_mask && sat3_mask) {
// Why would anyone do this?
ELOG("PREFIXD: Can't have both sat[0-1] and sat[-1-1] at the same time yet");
}
if (sat1_mask) {
if (sat1_mask != full_mask) {
ELOG("PREFIXD: Can't have partial sat1 mask yet (%i vs %i)", sat1_mask, full_mask);
}
if (IsD(dest.rd)) {
VMOV_immf(D0, 0.0);
VMOV_immf(D1, 1.0);
VMAX(F_32, dest.rd, dest.rd, D0);
VMIN(F_32, dest.rd, dest.rd, D1);
} else {
VMOV_immf(Q0, 1.0);
VMIN(F_32, dest.rd, dest.rd, Q0);
VMOV_immf(Q0, 0.0);
VMAX(F_32, dest.rd, dest.rd, Q0);
}
}
if (sat3_mask && sat1_mask != full_mask) {
if (sat3_mask != full_mask) {
ELOG("PREFIXD: Can't have partial sat3 mask yet (%i vs %i)", sat3_mask, full_mask);
}
if (IsD(dest.rd)) {
VMOV_immf(D0, 0.0);
VMOV_immf(D1, 1.0);
VMAX(F_32, dest.rd, dest.rd, D0);
VMIN(F_32, dest.rd, dest.rd, D1);
} else {
VMOV_immf(Q0, 1.0);
VMIN(F_32, dest.rd, dest.rd, Q0);
VMOV_immf(Q0, -1.0);
VMAX(F_32, dest.rd, dest.rd, Q0);
}
}
// Check for actual mask operation (unrelated to the "masks" above).
if (dest.backingRd != dest.rd) {
// This means that we need to apply the write mask, from rd to backingRd.
// What a pain. We can at least shortcut easy cases like half the register.
// And we can generate the masks easily with some of the crazy vector imm modes. (bits2bytes for example).
// So no need to load them from RAM.
int writeMask = (~(js.prefixD >> 8)) & 0xF;
if (writeMask == 3) {
ILOG("Doing writemask = 3");
VMOV(D_0(dest.rd), D_0(dest.backingRd));
} else {
// TODO
ELOG("PREFIXD: Arbitrary write masks not supported (%i / %i)", writeMask, full_mask);
VMOV(dest.backingRd, dest.rd);
}
}
void Jit::Comp_VecDo3(u32 op)
{
CONDITIONAL_DISABLE;
DISABLE;
// WARNING: No prefix support!
if (js.MayHavePrefix())
{
Comp_Generic(op);
js.EatPrefix();
return;
}
int vd = _VD;
int vs = _VS;
int vt = _VT;
void (ARMXEmitter::*triop)(ARMReg, ARMReg, ARMReg) = NULL;
switch (op >> 26)
{
case 24: //VFPU0
switch ((op >> 23)&7)
{
case 0: // d[i] = s[i] + t[i]; break; //vadd
triop = &ARMXEmitter::VADD;
break;
case 1: // d[i] = s[i] - t[i]; break; //vsub
triop = &ARMXEmitter::VSUB;
break;
case 7: // d[i] = s[i] / t[i]; break; //vdiv
triop = &ARMXEmitter::VDIV;
break;
}
break;
case 25: //VFPU1
switch ((op >> 23)&7)
{
case 0: // d[i] = s[i] * t[i]; break; //vmul
triop = &ARMXEmitter::VMUL;
break;
}
break;
}
if (!triop) {
DISABLE;
}
VectorSize sz = GetVecSize(op);
int n = GetNumVectorElements(sz);
u8 sregs[4], tregs[4], dregs[4];
GetVectorRegsPrefixS(sregs, sz, _VS);
GetVectorRegsPrefixT(tregs, sz, _VT);
GetVectorRegsPrefixD(dregs, sz, _VD);
MIPSReg tempregs[4];
for (int i = 0; i < n; i++) {
if (!IsOverlapSafeAllowS(dregs[i], i, n, sregs, n, tregs)) {
tempregs[i] = fpr.GetTempV();
} else {
fpr.MapRegV(dregs[i], (dregs[i] == sregs[i] || dregs[i] == tregs[i] ? 0 : MAP_NOINIT) | MAP_DIRTY);
tempregs[i] = dregs[i];
}
}
for (int i = 0; i < n; i++) {
fpr.SpillLockV(sregs[i]);
fpr.SpillLockV(tregs[i]);
fpr.MapRegV(sregs[i]);
fpr.MapRegV(tregs[i]);
fpr.MapRegV(tempregs[i]);
(this->*triop)(fpr.V(tempregs[i]), fpr.V(sregs[i]), fpr.V(tregs[i]));
fpr.ReleaseSpillLockV(sregs[i]);
fpr.ReleaseSpillLockV(tregs[i]);
}
fpr.MapRegsV(dregs, sz, MAP_DIRTY);
for (int i = 0; i < n; i++) {
if (dregs[i] != tempregs[i])
VMOV(fpr.V(dregs[i]), fpr.V(tempregs[i]));
}
ApplyPrefixD(dregs, sz);
fpr.ReleaseSpillLocks();
js.EatPrefix();
}