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;
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, 0.5, R0);
VABS(S1, fpr.V(vregs[i])); // S1 = fabs(x)
VSUB(S2, fpr.V(vregs[i]), S0); // S2 = fabs(x-0.5f) {VABD}
VABS(S2, S2);
VSUB(fpr.V(vregs[i]), S1, S2); // v[i] = S1 - S2 + 0.5f
VADD(fpr.V(vregs[i]), fpr.V(vregs[i]), S0);
} else if (sat == 3) {
// clamped = fabs(x) - fabs(x-1.0f); // [-1, 1]
fpr.MapRegV(vregs[i], MAP_DIRTY);
MOVI2F(S0, 1.0, R0);
VABS(S1, fpr.V(vregs[i])); // S1 = fabs(x)
VSUB(S2, fpr.V(vregs[i]), S0); // S2 = fabs(x-1.0f) {VABD}
VABS(S2, S2);
VSUB(fpr.V(vregs[i]), S1, S2); // v[i] = S1 - S2
}
}
}
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 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::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_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 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 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();
}