Merge pull request #1367 from unknownbrackets/jit-vfpu

Initial vscl, vmmul, vmscl, and v(h)tfm in x86 jit
This commit is contained in:
Henrik Rydgård committed 2013-04-20 05:55:21 -07:00
commit 8615de3f1b
5 files changed
+247 -21

No files matched your search

+16
View File
@@ -714,4 +714,20 @@ namespace MIPSComp
DISABLE;
}
void Jit::Comp_VScl(u32 op) {
DISABLE;
}
void Jit::Comp_Vmmul(u32 op) {
DISABLE;
}
void Jit::Comp_Vmscl(u32 op) {
DISABLE;
}
void Jit::Comp_Vtfm(u32 op) {
DISABLE;
}
}
+4
View File
@@ -193,6 +193,10 @@ public:
void Comp_Mftv(u32 op);
void Comp_Vmtvc(u32 op);
void Comp_Vmmov(u32 op);
void Comp_VScl(u32 op);
void Comp_Vmmul(u32 op);
void Comp_Vmscl(u32 op);
void Comp_Vtfm(u32 op);
ArmJitBlockCache *GetBlockCache() { return &blocks; }
+21 -21
View File
@@ -490,7 +490,7 @@ const MIPSInstruction tableVFPU1[8] =
{
INSTR("vmul",&Jit::Comp_VecDo3, Dis_VectorSet3, Int_VecDo3, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vdot",&Jit::Comp_VDot, Dis_VectorDot, Int_VDot, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vscl",&Jit::Comp_Generic, Dis_VScl, Int_VScl, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vscl",&Jit::Comp_VScl, Dis_VScl, Int_VScl, IS_VFPU|OUT_EAT_PREFIX),
{-2},
INSTR("vhdp",&Jit::Comp_Generic, Dis_Generic, Int_VHdp, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vcrs",&Jit::Comp_Generic, Dis_Vcrs, Int_Vcrs, IS_VFPU),
@@ -626,30 +626,30 @@ MIPSInstruction tableVFPU5[8] = //110111 xxx
const MIPSInstruction tableVFPU6[32] = //111100 xxx
{
//0
INSTR("vmmul",&Jit::Comp_Generic, Dis_MatrixMult, Int_Vmmul, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmmul",&Jit::Comp_Generic, Dis_MatrixMult, Int_Vmmul, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmmul",&Jit::Comp_Generic, Dis_MatrixMult, Int_Vmmul, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmmul",&Jit::Comp_Generic, Dis_MatrixMult, Int_Vmmul, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmmul",&Jit::Comp_Vmmul, Dis_MatrixMult, Int_Vmmul, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmmul",&Jit::Comp_Vmmul, Dis_MatrixMult, Int_Vmmul, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmmul",&Jit::Comp_Vmmul, Dis_MatrixMult, Int_Vmmul, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmmul",&Jit::Comp_Vmmul, Dis_MatrixMult, Int_Vmmul, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm2",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm2",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm2",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm2",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm2",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm2",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm2",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm2",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
//8
INSTR("v(h)tfm3",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm3",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm3",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm3",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm3",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm3",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm3",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm3",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm4",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm4",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm4",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm4",&Jit::Comp_Generic, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm4",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm4",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm4",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
INSTR("v(h)tfm4",&Jit::Comp_Vtfm, Dis_Vtfm, Int_Vtfm, IS_VFPU|OUT_EAT_PREFIX),
//16
INSTR("vmscl",&Jit::Comp_Generic, Dis_Generic, Int_Vmscl, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmscl",&Jit::Comp_Generic, Dis_Generic, Int_Vmscl, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmscl",&Jit::Comp_Generic, Dis_Generic, Int_Vmscl, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmscl",&Jit::Comp_Generic, Dis_Generic, Int_Vmscl, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmscl",&Jit::Comp_Vmscl, Dis_Generic, Int_Vmscl, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmscl",&Jit::Comp_Vmscl, Dis_Generic, Int_Vmscl, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmscl",&Jit::Comp_Vmscl, Dis_Generic, Int_Vmscl, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vmscl",&Jit::Comp_Vmscl, Dis_Generic, Int_Vmscl, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vcrsp.t/vqmul.q",&Jit::Comp_Generic, Dis_CrossQuat, Int_CrossQuat, IS_VFPU|OUT_EAT_PREFIX),
INSTR("vcrsp.t/vqmul.q",&Jit::Comp_Generic, Dis_CrossQuat, Int_CrossQuat, IS_VFPU|OUT_EAT_PREFIX),
+202
View File
@@ -732,4 +732,206 @@ void Jit::Comp_Vmmov(u32 op) {
fpr.ReleaseSpillLocks();
}
void Jit::Comp_VScl(u32 op) {
CONDITIONAL_DISABLE;
if (js.HasUnknownPrefix())
DISABLE;
VectorSize sz = GetVecSize(op);
int n = GetNumVectorElements(sz);
u8 sregs[4], dregs[4], scale;
GetVectorRegsPrefixS(sregs, sz, _VS);
// TODO: Prefixes seem strange...
GetVectorRegsPrefixT(&scale, V_Single, _VT);
GetVectorRegsPrefixD(dregs, sz, _VD);
// Move to XMM0 early, so we don't have to worry about overlap with scale.
MOVSS(XMM0, fpr.V(scale));
X64Reg 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.VX(reg);
}
else
{
fpr.MapRegV(dregs[i], (dregs[i] == sregs[i] ? 0 : MAP_NOINIT) | MAP_DIRTY);
fpr.SpillLockV(dregs[i]);
tempxregs[i] = fpr.VX(dregs[i]);
}
}
for (int i = 0; i < n; ++i)
{
if (!fpr.V(sregs[i]).IsSimpleReg(tempxregs[i]))
MOVSS(tempxregs[i], fpr.V(sregs[i]));
MULSS(tempxregs[i], R(XMM0));
}
for (int i = 0; i < n; ++i)
{
if (!fpr.V(dregs[i]).IsSimpleReg(tempxregs[i]))
MOVSS(fpr.V(dregs[i]), tempxregs[i]);
}
ApplyPrefixD(dregs, sz);
fpr.ReleaseSpillLocks();
}
void Jit::Comp_Vmmul(u32 op) {
CONDITIONAL_DISABLE;
// TODO: This probably ignores prefixes?
if (js.MayHavePrefix())
DISABLE;
MatrixSize sz = GetMtxSize(op);
int n = GetMatrixSide(sz);
u8 sregs[16], tregs[16], dregs[16];
GetMatrixRegs(sregs, sz, _VS);
GetMatrixRegs(tregs, sz, _VT);
GetMatrixRegs(dregs, sz, _VD);
// TODO: test overlap, fix non-optimal.
u8 tempregs[16];
for (int a = 0; a < n; a++)
{
for (int b = 0; b < n; b++)
{
XORPS(XMM0, R(XMM0));
for (int c = 0; c < n; c++)
{
MOVSS(XMM1, fpr.V(sregs[b * 4 + c]));
MULSS(XMM1, fpr.V(tregs[a * 4 + c]));
ADDSS(XMM0, R(XMM1));
}
u8 temp = (u8) fpr.GetTempV();
fpr.MapRegV(temp, MAP_NOINIT | MAP_DIRTY);
MOVSS(fpr.VX(temp), R(XMM0));
fpr.StoreFromRegisterV(temp);
tempregs[a * 4 + b] = temp;
}
}
for (int a = 0; a < n; a++)
{
for (int b = 0; b < n; b++)
{
u8 temp = tempregs[a * 4 + b];
fpr.MapRegV(temp, 0);
MOVSS(fpr.V(dregs[a * 4 + b]), fpr.VX(temp));
}
}
fpr.ReleaseSpillLocks();
}
void Jit::Comp_Vmscl(u32 op) {
CONDITIONAL_DISABLE;
// TODO: This probably ignores prefixes?
if (js.MayHavePrefix())
DISABLE;
MatrixSize sz = GetMtxSize(op);
int n = GetMatrixSide(sz);
u8 sregs[16], dregs[16], scale;
GetMatrixRegs(sregs, sz, _VS);
GetVectorRegs(&scale, V_Single, _VT);
GetMatrixRegs(dregs, sz, _VD);
// Move to XMM0 early, so we don't have to worry about overlap with scale.
MOVSS(XMM0, fpr.V(scale));
// TODO: test overlap, optimize.
u8 tempregs[16];
for (int a = 0; a < n; a++)
{
for (int b = 0; b < n; b++)
{
u8 temp = (u8) fpr.GetTempV();
fpr.MapRegV(temp, MAP_NOINIT | MAP_DIRTY);
MOVSS(fpr.VX(temp), fpr.V(sregs[a * 4 + b]));
MULSS(fpr.VX(temp), R(XMM0));
fpr.StoreFromRegisterV(temp);
tempregs[a * 4 + b] = temp;
}
}
for (int a = 0; a < n; a++)
{
for (int b = 0; b < n; b++)
{
u8 temp = tempregs[a * 4 + b];
fpr.MapRegV(temp, 0);
MOVSS(fpr.V(dregs[a * 4 + b]), fpr.VX(temp));
}
}
fpr.ReleaseSpillLocks();
}
void Jit::Comp_Vtfm(u32 op) {
CONDITIONAL_DISABLE;
// TODO: This probably ignores prefixes? Or maybe uses D?
if (js.MayHavePrefix())
DISABLE;
VectorSize sz = GetVecSize(op);
MatrixSize msz = GetMtxSize(op);
int n = GetNumVectorElements(sz);
int ins = (op >> 23) & 7;
bool homogenous = false;
if (n == ins)
{
n++;
sz = (VectorSize)((int)(sz) + 1);
msz = (MatrixSize)((int)(msz) + 1);
homogenous = true;
}
// Otherwise, n should already be ins + 1.
else if (n != ins + 1)
DISABLE;
u8 sregs[16], dregs[4], tregs[4];
GetMatrixRegs(sregs, msz, _VS);
GetVectorRegs(tregs, sz, _VT);
GetVectorRegs(dregs, sz, _VD);
// TODO: test overlap, optimize.
u8 tempregs[4];
for (int i = 0; i < n; i++)
{
XORPS(XMM0, R(XMM0));
for (int k = 0; k < n; k++)
{
MOVSS(XMM1, fpr.V(sregs[i * 4 + k]));
if (!homogenous || k != n - 1)
MULSS(XMM1, fpr.V(tregs[k]));
ADDSS(XMM0, R(XMM1));
}
u8 temp = (u8) fpr.GetTempV();
fpr.MapRegV(temp, MAP_NOINIT | MAP_DIRTY);
MOVSS(fpr.VX(temp), R(XMM0));
fpr.StoreFromRegisterV(temp);
tempregs[i] = temp;
}
for (int i = 0; i < n; i++)
{
u8 temp = tempregs[i];
fpr.MapRegV(temp, 0);
MOVSS(fpr.V(dregs[i]), fpr.VX(temp));
}
fpr.ReleaseSpillLocks();
}
}
+4
View File
@@ -203,6 +203,10 @@ public:
void Comp_Mftv(u32 op);
void Comp_Vmtvc(u32 op);
void Comp_Vmmov(u32 op);
void Comp_VScl(u32 op);
void Comp_Vmmul(u32 op);
void Comp_Vmscl(u32 op);
void Comp_Vtfm(u32 op);
void Comp_DoNothing(u32 op);