BT::RegisterCell BT::MachineEvaluator::eSUB(const RegisterCell &A1,
const RegisterCell &A2) const {
uint16_t W = A1.width();
assert(W == A2.width());
RegisterCell Res(W);
bool Borrow = false;
uint16_t I;
for (I = 0; I < W; ++I) {
const BitValue &V1 = A1[I];
const BitValue &V2 = A2[I];
if (!V1.num() || !V2.num())
break;
unsigned S = bool(V1) - bool(V2) - Borrow;
Res[I] = BitValue(S & 1);
Borrow = (S > 1);
}
for (; I < W; ++I) {
const BitValue &V1 = A1[I];
const BitValue &V2 = A2[I];
if (V1.is(Borrow)) {
Res[I] = BitValue::ref(V2);
break;
}
if (V2.is(Borrow))
Res[I] = BitValue::ref(V1);
else
break;
}
for (; I < W; ++I)
Res[I] = BitValue::self();
return Res;
}
BT::RegisterCell BT::MachineEvaluator::eADD(const RegisterCell &A1,
const RegisterCell &A2) const {
uint16_t W = A1.width();
assert(W == A2.width());
RegisterCell Res(W);
bool Carry = false;
uint16_t I;
for (I = 0; I < W; ++I) {
const BitValue &V1 = A1[I];
const BitValue &V2 = A2[I];
if (!V1.num() || !V2.num())
break;
unsigned S = bool(V1) + bool(V2) + Carry;
Res[I] = BitValue(S & 1);
Carry = (S > 1);
}
for (; I < W; ++I) {
const BitValue &V1 = A1[I];
const BitValue &V2 = A2[I];
// If the next bit is same as Carry, the result will be 0 plus the
// other bit. The Carry bit will remain unchanged.
if (V1.is(Carry))
Res[I] = BitValue::ref(V2);
else if (V2.is(Carry))
Res[I] = BitValue::ref(V1);
else
break;
}
for (; I < W; ++I)
Res[I] = BitValue::self();
return Res;
}
BT::RegisterCell BT::MachineEvaluator::eASL(const RegisterCell &A1,
uint16_t Sh) const {
assert(Sh <= A1.width());
RegisterCell Res = RegisterCell::ref(A1);
Res.rol(Sh);
Res.fill(0, Sh, BitValue::Zero);
return Res;
}
BT::RegisterCell BT::MachineEvaluator::eZXT(const RegisterCell &A1,
uint16_t FromN) const {
uint16_t W = A1.width();
assert(FromN <= W);
RegisterCell Res = RegisterCell::ref(A1);
Res.fill(FromN, W, BitValue::Zero);
return Res;
}
BT::RegisterCell BT::MachineEvaluator::eMLU(const RegisterCell &A1,
const RegisterCell &A2) const {
uint16_t W = A1.width() + A2.width();
uint16_t Z = A1.ct(false) + A2.ct(false);
RegisterCell Res(W);
Res.fill(0, Z, BitValue::Zero);
Res.fill(Z, W, BitValue::self());
return Res;
}
BT::RegisterCell BT::MachineEvaluator::eCTB(const RegisterCell &A1, bool B,
uint16_t W) const {
uint16_t C = A1.ct(B), AW = A1.width();
// If the last trailing non-B bit is not a constant, then we don't know
// the real count.
if ((C < AW && A1[C].num()) || C == AW)
return eIMM(C, W);
return RegisterCell::self(0, W);
}
BT::RegisterCell BT::MachineEvaluator::eLSR(const RegisterCell &A1,
uint16_t Sh) const {
uint16_t W = A1.width();
assert(Sh <= W);
RegisterCell Res = RegisterCell::ref(A1);
Res.rol(W-Sh);
Res.fill(W-Sh, W, BitValue::Zero);
return Res;
}
BT::RegisterCell BT::MachineEvaluator::eSXT(const RegisterCell &A1,
uint16_t FromN) const {
uint16_t W = A1.width();
assert(FromN <= W);
RegisterCell Res = RegisterCell::ref(A1);
BitValue Sign = Res[FromN-1];
// Sign-extend "inreg".
Res.fill(FromN, W, Sign);
return Res;
}
BT::RegisterCell BT::MachineEvaluator::eINS(const RegisterCell &A1,
const RegisterCell &A2, uint16_t AtN) const {
uint16_t W1 = A1.width(), W2 = A2.width();
(void)W1;
assert(AtN < W1 && AtN+W2 <= W1);
// Copy bits from A1, insert A2 at position AtN.
RegisterCell Res = RegisterCell::ref(A1);
if (W2 > 0)
Res.insert(RegisterCell::ref(A2), BT::BitMask(AtN, AtN+W2-1));
return Res;
}
BT::RegisterCell BT::MachineEvaluator::eCLR(const RegisterCell &A1,
uint16_t BitN) const {
assert(BitN < A1.width());
RegisterCell Res = RegisterCell::ref(A1);
Res[BitN] = BitValue::Zero;
return Res;
}
// Check if the cell represents a compile-time integer value.
bool BT::MachineEvaluator::isInt(const RegisterCell &A) const {
uint16_t W = A.width();
for (uint16_t i = 0; i < W; ++i)
if (!A[i].is(0) && !A[i].is(1))
return false;
return true;
}
void BT::visitPHI(const MachineInstr &PI) {
int ThisN = PI.getParent()->getNumber();
if (Trace)
dbgs() << "Visit FI(" << printMBBReference(*PI.getParent()) << "): " << PI;
const MachineOperand &MD = PI.getOperand(0);
assert(MD.getSubReg() == 0 && "Unexpected sub-register in definition");
RegisterRef DefRR(MD);
uint16_t DefBW = ME.getRegBitWidth(DefRR);
RegisterCell DefC = ME.getCell(DefRR, Map);
if (DefC == RegisterCell::self(DefRR.Reg, DefBW)) // XXX slow
return;
bool Changed = false;
for (unsigned i = 1, n = PI.getNumOperands(); i < n; i += 2) {
const MachineBasicBlock *PB = PI.getOperand(i + 1).getMBB();
int PredN = PB->getNumber();
if (Trace)
dbgs() << " edge " << printMBBReference(*PB) << "->"
<< printMBBReference(*PI.getParent());
if (!EdgeExec.count(CFGEdge(PredN, ThisN))) {
if (Trace)
dbgs() << " not executable\n";
continue;
}
RegisterRef RU = PI.getOperand(i);
RegisterCell ResC = ME.getCell(RU, Map);
if (Trace)
dbgs() << " input reg: " << printReg(RU.Reg, &ME.TRI, RU.Sub)
<< " cell: " << ResC << "\n";
Changed |= DefC.meet(ResC, DefRR.Reg);
}
if (Changed) {
if (Trace)
dbgs() << "Output: " << printReg(DefRR.Reg, &ME.TRI, DefRR.Sub)
<< " cell: " << DefC << "\n";
ME.putCell(DefRR, DefC, Map);
visitUsesOf(DefRR.Reg);
}
}
BT::RegisterCell BT::MachineEvaluator::eXOR(const RegisterCell &A1,
const RegisterCell &A2) const {
uint16_t W = A1.width();
assert(W == A2.width());
RegisterCell Res(W);
for (uint16_t i = 0; i < W; ++i) {
const BitValue &V1 = A1[i];
const BitValue &V2 = A2[i];
if (V1.is(0))
Res[i] = BitValue::ref(V2);
else if (V2.is(0))
Res[i] = BitValue::ref(V1);
else if (V1 == V2)
Res[i] = BitValue::Zero;
else
Res[i] = BitValue::self();
}
return Res;
}
// Convert a cell to the integer value. The result must fit in uint64_t.
uint64_t BT::MachineEvaluator::toInt(const RegisterCell &A) const {
assert(isInt(A));
uint64_t Val = 0;
uint16_t W = A.width();
for (uint16_t i = 0; i < W; ++i) {
Val <<= 1;
Val |= A[i].is(1);
}
return Val;
}
void BT::MachineEvaluator::putCell(const RegisterRef &RR, RegisterCell RC,
CellMapType &M) const {
// While updating the cell map can be done in a meaningful way for
// a part of a register, it makes little sense to implement it as the
// SSA representation would never contain such "partial definitions".
if (!TargetRegisterInfo::isVirtualRegister(RR.Reg))
return;
assert(RR.Sub == 0 && "Unexpected sub-register in definition");
// Eliminate all ref-to-reg-0 bit values: replace them with "self".
M[RR.Reg] = RC.regify(RR.Reg);
}
BT::RegisterCell BT::MachineEvaluator::eXTR(const RegisterCell &A1,
uint16_t B, uint16_t E) const {
uint16_t W = A1.width();
assert(B < W && E <= W);
if (B == E)
return RegisterCell(0);
uint16_t Last = (E > 0) ? E-1 : W-1;
RegisterCell Res = RegisterCell::ref(A1).extract(BT::BitMask(B, Last));
// Return shorter cell.
return Res;
}
BT::RegisterCell BT::MachineEvaluator::eNOT(const RegisterCell &A1) const {
uint16_t W = A1.width();
RegisterCell Res(W);
for (uint16_t i = 0; i < W; ++i) {
const BitValue &V = A1[i];
if (V.is(0))
Res[i] = BitValue::One;
else if (V.is(1))
Res[i] = BitValue::Zero;
else
Res[i] = BitValue::self();
}
return Res;
}
void BT::MachineEvaluator::putCell(const RegisterRef &RR, RegisterCell RC,
CellMapType &M) const {
// While updating the cell map can be done in a meaningful way for
// a part of a register, it makes little sense to implement it as the
// SSA representation would never contain such "partial definitions".
if (!TargetRegisterInfo::isVirtualRegister(RR.Reg))
return;
assert(RR.Sub == 0 && "Unexpected sub-register in definition");
// Eliminate all ref-to-reg-0 bit values: replace them with "self".
for (unsigned i = 0, n = RC.width(); i < n; ++i) {
const BitValue &V = RC[i];
if (V.Type == BitValue::Ref && V.RefI.Reg == 0)
RC[i].RefI = BitRef(RR.Reg, i);
}
M[RR.Reg] = RC;
}
void BT::visitNonBranch(const MachineInstr &MI) {
if (Trace) {
int ThisN = MI.getParent()->getNumber();
dbgs() << "Visit MI(BB#" << ThisN << "): " << MI;
}
if (MI.isDebugValue())
return;
assert(!MI.isBranch() && "Unexpected branch instruction");
CellMapType ResMap;
bool Eval = ME.evaluate(MI, Map, ResMap);
if (Trace && Eval) {
for (unsigned i = 0, n = MI.getNumOperands(); i < n; ++i) {
const MachineOperand &MO = MI.getOperand(i);
if (!MO.isReg() || !MO.isUse())
continue;
RegisterRef RU(MO);
dbgs() << " input reg: " << PrintReg(RU.Reg, &ME.TRI, RU.Sub)
<< " cell: " << ME.getCell(RU, Map) << "\n";
}
dbgs() << "Outputs:\n";
for (CellMapType::iterator I = ResMap.begin(), E = ResMap.end();
I != E; ++I) {
RegisterRef RD(I->first);
dbgs() << " " << PrintReg(I->first, &ME.TRI) << " cell: "
<< ME.getCell(RD, ResMap) << "\n";
}
}
// Iterate over all definitions of the instruction, and update the
// cells accordingly.
for (unsigned i = 0, n = MI.getNumOperands(); i < n; ++i) {
const MachineOperand &MO = MI.getOperand(i);
// Visit register defs only.
if (!MO.isReg() || !MO.isDef())
continue;
RegisterRef RD(MO);
assert(RD.Sub == 0 && "Unexpected sub-register in definition");
if (!TargetRegisterInfo::isVirtualRegister(RD.Reg))
continue;
bool Changed = false;
if (!Eval || ResMap.count(RD.Reg) == 0) {
// Set to "ref" (aka "bottom").
uint16_t DefBW = ME.getRegBitWidth(RD);
RegisterCell RefC = RegisterCell::self(RD.Reg, DefBW);
if (RefC != ME.getCell(RD, Map)) {
ME.putCell(RD, RefC, Map);
Changed = true;
}
} else {
RegisterCell DefC = ME.getCell(RD, Map);
RegisterCell ResC = ME.getCell(RD, ResMap);
// This is a non-phi instruction, so the values of the inputs come
// from the same registers each time this instruction is evaluated.
// During the propagation, the values of the inputs can become lowered
// in the sense of the lattice operation, which may cause different
// results to be calculated in subsequent evaluations. This should
// not cause the bottoming of the result in the map, since the new
// result is already reflecting the lowered inputs.
for (uint16_t i = 0, w = DefC.width(); i < w; ++i) {
BitValue &V = DefC[i];
// Bits that are already "bottom" should not be updated.
if (V.Type == BitValue::Ref && V.RefI.Reg == RD.Reg)
continue;
// Same for those that are identical in DefC and ResC.
if (V == ResC[i])
continue;
V = ResC[i];
Changed = true;
}
if (Changed)
ME.putCell(RD, DefC, Map);
}
if (Changed)
visitUsesOf(RD.Reg);
}
}
bool HexagonEvaluator::evaluate(const MachineInstr *MI,
const CellMapType &Inputs, CellMapType &Outputs) const {
unsigned NumDefs = 0;
// Sanity verification: there should not be any defs with subregisters.
for (unsigned i = 0, n = MI->getNumOperands(); i < n; ++i) {
const MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg() || !MO.isDef())
continue;
NumDefs++;
assert(MO.getSubReg() == 0);
}
if (NumDefs == 0)
return false;
if (MI->mayLoad())
return evaluateLoad(MI, Inputs, Outputs);
// Check COPY instructions that copy formal parameters into virtual
// registers. Such parameters can be sign- or zero-extended at the
// call site, and we should take advantage of this knowledge. The MRI
// keeps a list of pairs of live-in physical and virtual registers,
// which provides information about which virtual registers will hold
// the argument values. The function will still contain instructions
// defining those virtual registers, and in practice those are COPY
// instructions from a physical to a virtual register. In such cases,
// applying the argument extension to the virtual register can be seen
// as simply mirroring the extension that had already been applied to
// the physical register at the call site. If the defining instruction
// was not a COPY, it would not be clear how to mirror that extension
// on the callee's side. For that reason, only check COPY instructions
// for potential extensions.
if (MI->isCopy()) {
if (evaluateFormalCopy(MI, Inputs, Outputs))
return true;
}
// Beyond this point, if any operand is a global, skip that instruction.
// The reason is that certain instructions that can take an immediate
// operand can also have a global symbol in that operand. To avoid
// checking what kind of operand a given instruction has individually
// for each instruction, do it here. Global symbols as operands gene-
// rally do not provide any useful information.
for (unsigned i = 0, n = MI->getNumOperands(); i < n; ++i) {
const MachineOperand &MO = MI->getOperand(i);
if (MO.isGlobal() || MO.isBlockAddress() || MO.isSymbol() || MO.isJTI() ||
MO.isCPI())
return false;
}
RegisterRefs Reg(MI);
unsigned Opc = MI->getOpcode();
using namespace Hexagon;
#define op(i) MI->getOperand(i)
#define rc(i) RegisterCell::ref(getCell(Reg[i],Inputs))
#define im(i) MI->getOperand(i).getImm()
// If the instruction has no register operands, skip it.
if (Reg.size() == 0)
return false;
// Record result for register in operand 0.
auto rr0 = [this,Reg] (const BT::RegisterCell &Val, CellMapType &Outputs)
-> bool {
putCell(Reg[0], Val, Outputs);
return true;
};
// Get the cell corresponding to the N-th operand.
auto cop = [this,Reg,MI,Inputs] (unsigned N, uint16_t W)
-> BT::RegisterCell {
const MachineOperand &Op = MI->getOperand(N);
if (Op.isImm())
return eIMM(Op.getImm(), W);
if (!Op.isReg())
return RegisterCell::self(0, W);
assert(getRegBitWidth(Reg[N]) == W && "Register width mismatch");
return rc(N);
};
// Extract RW low bits of the cell.
auto lo = [this] (const BT::RegisterCell &RC, uint16_t RW)
-> BT::RegisterCell {
assert(RW <= RC.width());
return eXTR(RC, 0, RW);
};
// Extract RW high bits of the cell.
auto hi = [this] (const BT::RegisterCell &RC, uint16_t RW)
-> BT::RegisterCell {
uint16_t W = RC.width();
assert(RW <= W);
return eXTR(RC, W-RW, W);
};
// Extract N-th halfword (counting from the least significant position).
auto half = [this] (const BT::RegisterCell &RC, unsigned N)
-> BT::RegisterCell {
assert(N*16+16 <= RC.width());
return eXTR(RC, N*16, N*16+16);
};
// Shuffle bits (pick even/odd from cells and merge into result).
auto shuffle = [this] (const BT::RegisterCell &Rs, const BT::RegisterCell &Rt,
uint16_t BW, bool Odd) -> BT::RegisterCell {
uint16_t I = Odd, Ws = Rs.width();
assert(Ws == Rt.width());
RegisterCell RC = eXTR(Rt, I*BW, I*BW+BW).cat(eXTR(Rs, I*BW, I*BW+BW));
I += 2;
while (I*BW < Ws) {
RC.cat(eXTR(Rt, I*BW, I*BW+BW)).cat(eXTR(Rs, I*BW, I*BW+BW));
I += 2;
}
return RC;
};
// The bitwidth of the 0th operand. In most (if not all) of the
// instructions below, the 0th operand is the defined register.
// Pre-compute the bitwidth here, because it is needed in many cases
// cases below.
uint16_t W0 = (Reg[0].Reg != 0) ? getRegBitWidth(Reg[0]) : 0;
switch (Opc) {
// Transfer immediate:
case A2_tfrsi:
case A2_tfrpi:
case CONST32:
case CONST32_Float_Real:
case CONST32_Int_Real:
case CONST64_Float_Real:
case CONST64_Int_Real:
return rr0(eIMM(im(1), W0), Outputs);
case TFR_PdFalse:
return rr0(RegisterCell(W0).fill(0, W0, BT::BitValue::Zero), Outputs);
case TFR_PdTrue:
return rr0(RegisterCell(W0).fill(0, W0, BT::BitValue::One), Outputs);
case TFR_FI: {
int FI = op(1).getIndex();
int Off = op(2).getImm();
unsigned A = MFI.getObjectAlignment(FI) + std::abs(Off);
unsigned L = Log2_32(A);
RegisterCell RC = RegisterCell::self(Reg[0].Reg, W0);
RC.fill(0, L, BT::BitValue::Zero);
return rr0(RC, Outputs);
}
// Transfer register:
case A2_tfr:
case A2_tfrp:
case C2_pxfer_map:
return rr0(rc(1), Outputs);
case C2_tfrpr: {
uint16_t RW = W0;
uint16_t PW = 8; // XXX Pred size: getRegBitWidth(Reg[1]);
assert(PW <= RW);
RegisterCell PC = eXTR(rc(1), 0, PW);
RegisterCell RC = RegisterCell(RW).insert(PC, BT::BitMask(0, PW-1));
RC.fill(PW, RW, BT::BitValue::Zero);
return rr0(RC, Outputs);
}
case C2_tfrrp: {
RegisterCell RC = RegisterCell::self(Reg[0].Reg, W0);
W0 = 8; // XXX Pred size
return rr0(eINS(RC, eXTR(rc(1), 0, W0), 0), Outputs);
}
// Arithmetic:
case A2_abs:
case A2_absp:
// TODO
break;
case A2_addsp: {
uint16_t W1 = getRegBitWidth(Reg[1]);
assert(W0 == 64 && W1 == 32);
RegisterCell CW = RegisterCell(W0).insert(rc(1), BT::BitMask(0, W1-1));
RegisterCell RC = eADD(eSXT(CW, W1), rc(2));
return rr0(RC, Outputs);
}
case A2_add:
case A2_addp:
return rr0(eADD(rc(1), rc(2)), Outputs);
case A2_addi:
return rr0(eADD(rc(1), eIMM(im(2), W0)), Outputs);
case S4_addi_asl_ri: {
RegisterCell RC = eADD(eIMM(im(1), W0), eASL(rc(2), im(3)));
return rr0(RC, Outputs);
}
case S4_addi_lsr_ri: {
RegisterCell RC = eADD(eIMM(im(1), W0), eLSR(rc(2), im(3)));
return rr0(RC, Outputs);
}
case S4_addaddi: {
RegisterCell RC = eADD(rc(1), eADD(rc(2), eIMM(im(3), W0)));
return rr0(RC, Outputs);
}
case M4_mpyri_addi: {
RegisterCell M = eMLS(rc(2), eIMM(im(3), W0));
RegisterCell RC = eADD(eIMM(im(1), W0), lo(M, W0));
return rr0(RC, Outputs);
}
case M4_mpyrr_addi: {
RegisterCell M = eMLS(rc(2), rc(3));
RegisterCell RC = eADD(eIMM(im(1), W0), lo(M, W0));
return rr0(RC, Outputs);
}
case M4_mpyri_addr_u2: {
RegisterCell M = eMLS(eIMM(im(2), W0), rc(3));
RegisterCell RC = eADD(rc(1), lo(M, W0));
return rr0(RC, Outputs);
}
case M4_mpyri_addr: {
RegisterCell M = eMLS(rc(2), eIMM(im(3), W0));
RegisterCell RC = eADD(rc(1), lo(M, W0));
return rr0(RC, Outputs);
}
case M4_mpyrr_addr: {
RegisterCell M = eMLS(rc(2), rc(3));
RegisterCell RC = eADD(rc(1), lo(M, W0));
return rr0(RC, Outputs);
}
case S4_subaddi: {
RegisterCell RC = eADD(rc(1), eSUB(eIMM(im(2), W0), rc(3)));
return rr0(RC, Outputs);
}
case M2_accii: {
RegisterCell RC = eADD(rc(1), eADD(rc(2), eIMM(im(3), W0)));
return rr0(RC, Outputs);
}
case M2_acci: {
RegisterCell RC = eADD(rc(1), eADD(rc(2), rc(3)));
return rr0(RC, Outputs);
}
case M2_subacc: {
RegisterCell RC = eADD(rc(1), eSUB(rc(2), rc(3)));
return rr0(RC, Outputs);
}
case S2_addasl_rrri: {
RegisterCell RC = eADD(rc(1), eASL(rc(2), im(3)));
return rr0(RC, Outputs);
}
case C4_addipc: {
RegisterCell RPC = RegisterCell::self(Reg[0].Reg, W0);
RPC.fill(0, 2, BT::BitValue::Zero);
return rr0(eADD(RPC, eIMM(im(2), W0)), Outputs);
}
case A2_sub:
case A2_subp:
return rr0(eSUB(rc(1), rc(2)), Outputs);
case A2_subri:
return rr0(eSUB(eIMM(im(1), W0), rc(2)), Outputs);
case S4_subi_asl_ri: {
RegisterCell RC = eSUB(eIMM(im(1), W0), eASL(rc(2), im(3)));
return rr0(RC, Outputs);
}
case S4_subi_lsr_ri: {
RegisterCell RC = eSUB(eIMM(im(1), W0), eLSR(rc(2), im(3)));
return rr0(RC, Outputs);
}
case M2_naccii: {
RegisterCell RC = eSUB(rc(1), eADD(rc(2), eIMM(im(3), W0)));
return rr0(RC, Outputs);
}
case M2_nacci: {
RegisterCell RC = eSUB(rc(1), eADD(rc(2), rc(3)));
return rr0(RC, Outputs);
}
// 32-bit negation is done by "Rd = A2_subri 0, Rs"
case A2_negp:
return rr0(eSUB(eIMM(0, W0), rc(1)), Outputs);
case M2_mpy_up: {
RegisterCell M = eMLS(rc(1), rc(2));
return rr0(hi(M, W0), Outputs);
}
case M2_dpmpyss_s0:
return rr0(eMLS(rc(1), rc(2)), Outputs);
case M2_dpmpyss_acc_s0:
return rr0(eADD(rc(1), eMLS(rc(2), rc(3))), Outputs);
case M2_dpmpyss_nac_s0:
return rr0(eSUB(rc(1), eMLS(rc(2), rc(3))), Outputs);
case M2_mpyi: {
RegisterCell M = eMLS(rc(1), rc(2));
return rr0(lo(M, W0), Outputs);
}
case M2_macsip: {
RegisterCell M = eMLS(rc(2), eIMM(im(3), W0));
RegisterCell RC = eADD(rc(1), lo(M, W0));
return rr0(RC, Outputs);
}
case M2_macsin: {
RegisterCell M = eMLS(rc(2), eIMM(im(3), W0));
RegisterCell RC = eSUB(rc(1), lo(M, W0));
return rr0(RC, Outputs);
}
case M2_maci: {
RegisterCell M = eMLS(rc(2), rc(3));
RegisterCell RC = eADD(rc(1), lo(M, W0));
return rr0(RC, Outputs);
}
case M2_mpysmi: {
RegisterCell M = eMLS(rc(1), eIMM(im(2), W0));
return rr0(lo(M, 32), Outputs);
}
case M2_mpysin: {
RegisterCell M = eMLS(rc(1), eIMM(-im(2), W0));
return rr0(lo(M, 32), Outputs);
}
case M2_mpysip: {
RegisterCell M = eMLS(rc(1), eIMM(im(2), W0));
return rr0(lo(M, 32), Outputs);
}
case M2_mpyu_up: {
RegisterCell M = eMLU(rc(1), rc(2));
return rr0(hi(M, W0), Outputs);
}
case M2_dpmpyuu_s0:
return rr0(eMLU(rc(1), rc(2)), Outputs);
case M2_dpmpyuu_acc_s0:
return rr0(eADD(rc(1), eMLU(rc(2), rc(3))), Outputs);
case M2_dpmpyuu_nac_s0:
return rr0(eSUB(rc(1), eMLU(rc(2), rc(3))), Outputs);
//case M2_mpysu_up:
// Logical/bitwise:
case A2_andir:
return rr0(eAND(rc(1), eIMM(im(2), W0)), Outputs);
case A2_and:
case A2_andp:
return rr0(eAND(rc(1), rc(2)), Outputs);
case A4_andn:
case A4_andnp:
return rr0(eAND(rc(1), eNOT(rc(2))), Outputs);
case S4_andi_asl_ri: {
RegisterCell RC = eAND(eIMM(im(1), W0), eASL(rc(2), im(3)));
return rr0(RC, Outputs);
}
case S4_andi_lsr_ri: {
RegisterCell RC = eAND(eIMM(im(1), W0), eLSR(rc(2), im(3)));
return rr0(RC, Outputs);
}
case M4_and_and:
return rr0(eAND(rc(1), eAND(rc(2), rc(3))), Outputs);
case M4_and_andn:
return rr0(eAND(rc(1), eAND(rc(2), eNOT(rc(3)))), Outputs);
case M4_and_or:
return rr0(eAND(rc(1), eORL(rc(2), rc(3))), Outputs);
case M4_and_xor:
return rr0(eAND(rc(1), eXOR(rc(2), rc(3))), Outputs);
case A2_orir:
return rr0(eORL(rc(1), eIMM(im(2), W0)), Outputs);
case A2_or:
case A2_orp:
return rr0(eORL(rc(1), rc(2)), Outputs);
case A4_orn:
case A4_ornp:
return rr0(eORL(rc(1), eNOT(rc(2))), Outputs);
case S4_ori_asl_ri: {
RegisterCell RC = eORL(eIMM(im(1), W0), eASL(rc(2), im(3)));
return rr0(RC, Outputs);
}
case S4_ori_lsr_ri: {
RegisterCell RC = eORL(eIMM(im(1), W0), eLSR(rc(2), im(3)));
return rr0(RC, Outputs);
}
case M4_or_and:
return rr0(eORL(rc(1), eAND(rc(2), rc(3))), Outputs);
case M4_or_andn:
return rr0(eORL(rc(1), eAND(rc(2), eNOT(rc(3)))), Outputs);
case S4_or_andi:
case S4_or_andix: {
RegisterCell RC = eORL(rc(1), eAND(rc(2), eIMM(im(3), W0)));
return rr0(RC, Outputs);
}
case S4_or_ori: {
RegisterCell RC = eORL(rc(1), eORL(rc(2), eIMM(im(3), W0)));
return rr0(RC, Outputs);
}
case M4_or_or:
return rr0(eORL(rc(1), eORL(rc(2), rc(3))), Outputs);
case M4_or_xor:
return rr0(eORL(rc(1), eXOR(rc(2), rc(3))), Outputs);
case A2_xor:
case A2_xorp:
return rr0(eXOR(rc(1), rc(2)), Outputs);
case M4_xor_and:
return rr0(eXOR(rc(1), eAND(rc(2), rc(3))), Outputs);
case M4_xor_andn:
return rr0(eXOR(rc(1), eAND(rc(2), eNOT(rc(3)))), Outputs);
case M4_xor_or:
return rr0(eXOR(rc(1), eORL(rc(2), rc(3))), Outputs);
case M4_xor_xacc:
return rr0(eXOR(rc(1), eXOR(rc(2), rc(3))), Outputs);
case A2_not:
case A2_notp:
return rr0(eNOT(rc(1)), Outputs);
case S2_asl_i_r:
case S2_asl_i_p:
return rr0(eASL(rc(1), im(2)), Outputs);
case A2_aslh:
return rr0(eASL(rc(1), 16), Outputs);
case S2_asl_i_r_acc:
case S2_asl_i_p_acc:
return rr0(eADD(rc(1), eASL(rc(2), im(3))), Outputs);
case S2_asl_i_r_nac:
case S2_asl_i_p_nac:
return rr0(eSUB(rc(1), eASL(rc(2), im(3))), Outputs);
case S2_asl_i_r_and:
case S2_asl_i_p_and:
return rr0(eAND(rc(1), eASL(rc(2), im(3))), Outputs);
case S2_asl_i_r_or:
case S2_asl_i_p_or:
return rr0(eORL(rc(1), eASL(rc(2), im(3))), Outputs);
case S2_asl_i_r_xacc:
case S2_asl_i_p_xacc:
return rr0(eXOR(rc(1), eASL(rc(2), im(3))), Outputs);
case S2_asl_i_vh:
case S2_asl_i_vw:
// TODO
break;
case S2_asr_i_r:
case S2_asr_i_p:
return rr0(eASR(rc(1), im(2)), Outputs);
case A2_asrh:
return rr0(eASR(rc(1), 16), Outputs);
case S2_asr_i_r_acc:
case S2_asr_i_p_acc:
return rr0(eADD(rc(1), eASR(rc(2), im(3))), Outputs);
case S2_asr_i_r_nac:
case S2_asr_i_p_nac:
return rr0(eSUB(rc(1), eASR(rc(2), im(3))), Outputs);
case S2_asr_i_r_and:
case S2_asr_i_p_and:
return rr0(eAND(rc(1), eASR(rc(2), im(3))), Outputs);
case S2_asr_i_r_or:
case S2_asr_i_p_or:
return rr0(eORL(rc(1), eASR(rc(2), im(3))), Outputs);
case S2_asr_i_r_rnd: {
// The input is first sign-extended to 64 bits, then the output
// is truncated back to 32 bits.
assert(W0 == 32);
RegisterCell XC = eSXT(rc(1).cat(eIMM(0, W0)), W0);
RegisterCell RC = eASR(eADD(eASR(XC, im(2)), eIMM(1, 2*W0)), 1);
return rr0(eXTR(RC, 0, W0), Outputs);
}
case S2_asr_i_r_rnd_goodsyntax: {
int64_t S = im(2);
if (S == 0)
return rr0(rc(1), Outputs);
// Result: S2_asr_i_r_rnd Rs, u5-1
RegisterCell XC = eSXT(rc(1).cat(eIMM(0, W0)), W0);
RegisterCell RC = eLSR(eADD(eASR(XC, S-1), eIMM(1, 2*W0)), 1);
return rr0(eXTR(RC, 0, W0), Outputs);
}
case S2_asr_r_vh:
case S2_asr_i_vw:
case S2_asr_i_svw_trun:
// TODO
break;
case S2_lsr_i_r:
case S2_lsr_i_p:
return rr0(eLSR(rc(1), im(2)), Outputs);
case S2_lsr_i_r_acc:
case S2_lsr_i_p_acc:
return rr0(eADD(rc(1), eLSR(rc(2), im(3))), Outputs);
case S2_lsr_i_r_nac:
case S2_lsr_i_p_nac:
return rr0(eSUB(rc(1), eLSR(rc(2), im(3))), Outputs);
case S2_lsr_i_r_and:
case S2_lsr_i_p_and:
return rr0(eAND(rc(1), eLSR(rc(2), im(3))), Outputs);
case S2_lsr_i_r_or:
case S2_lsr_i_p_or:
return rr0(eORL(rc(1), eLSR(rc(2), im(3))), Outputs);
case S2_lsr_i_r_xacc:
case S2_lsr_i_p_xacc:
return rr0(eXOR(rc(1), eLSR(rc(2), im(3))), Outputs);
case S2_clrbit_i: {
RegisterCell RC = rc(1);
RC[im(2)] = BT::BitValue::Zero;
return rr0(RC, Outputs);
}
case S2_setbit_i: {
RegisterCell RC = rc(1);
RC[im(2)] = BT::BitValue::One;
return rr0(RC, Outputs);
}
case S2_togglebit_i: {
RegisterCell RC = rc(1);
uint16_t BX = im(2);
RC[BX] = RC[BX].is(0) ? BT::BitValue::One
: RC[BX].is(1) ? BT::BitValue::Zero
: BT::BitValue::self();
return rr0(RC, Outputs);
}
case A4_bitspliti: {
uint16_t W1 = getRegBitWidth(Reg[1]);
uint16_t BX = im(2);
// Res.uw[1] = Rs[bx+1:], Res.uw[0] = Rs[0:bx]
const BT::BitValue Zero = BT::BitValue::Zero;
RegisterCell RZ = RegisterCell(W0).fill(BX, W1, Zero)
.fill(W1+(W1-BX), W0, Zero);
RegisterCell BF1 = eXTR(rc(1), 0, BX), BF2 = eXTR(rc(1), BX, W1);
RegisterCell RC = eINS(eINS(RZ, BF1, 0), BF2, W1);
return rr0(RC, Outputs);
}
case S4_extract:
case S4_extractp:
case S2_extractu:
case S2_extractup: {
uint16_t Wd = im(2), Of = im(3);
assert(Wd <= W0);
if (Wd == 0)
return rr0(eIMM(0, W0), Outputs);
// If the width extends beyond the register size, pad the register
// with 0 bits.
RegisterCell Pad = (Wd+Of > W0) ? rc(1).cat(eIMM(0, Wd+Of-W0)) : rc(1);
RegisterCell Ext = eXTR(Pad, Of, Wd+Of);
// Ext is short, need to extend it with 0s or sign bit.
RegisterCell RC = RegisterCell(W0).insert(Ext, BT::BitMask(0, Wd-1));
if (Opc == S2_extractu || Opc == S2_extractup)
return rr0(eZXT(RC, Wd), Outputs);
return rr0(eSXT(RC, Wd), Outputs);
}
case S2_insert:
case S2_insertp: {
uint16_t Wd = im(3), Of = im(4);
assert(Wd < W0 && Of < W0);
// If Wd+Of exceeds W0, the inserted bits are truncated.
if (Wd+Of > W0)
Wd = W0-Of;
if (Wd == 0)
return rr0(rc(1), Outputs);
return rr0(eINS(rc(1), eXTR(rc(2), 0, Wd), Of), Outputs);
}
// Bit permutations:
case A2_combineii:
case A4_combineii:
case A4_combineir:
case A4_combineri:
case A2_combinew:
assert(W0 % 2 == 0);
return rr0(cop(2, W0/2).cat(cop(1, W0/2)), Outputs);
case A2_combine_ll:
case A2_combine_lh:
case A2_combine_hl:
case A2_combine_hh: {
assert(W0 == 32);
assert(getRegBitWidth(Reg[1]) == 32 && getRegBitWidth(Reg[2]) == 32);
// Low half in the output is 0 for _ll and _hl, 1 otherwise:
unsigned LoH = !(Opc == A2_combine_ll || Opc == A2_combine_hl);
// High half in the output is 0 for _ll and _lh, 1 otherwise:
unsigned HiH = !(Opc == A2_combine_ll || Opc == A2_combine_lh);
RegisterCell R1 = rc(1);
RegisterCell R2 = rc(2);
RegisterCell RC = half(R2, LoH).cat(half(R1, HiH));
return rr0(RC, Outputs);
}
case S2_packhl: {
assert(W0 == 64);
assert(getRegBitWidth(Reg[1]) == 32 && getRegBitWidth(Reg[2]) == 32);
RegisterCell R1 = rc(1);
RegisterCell R2 = rc(2);
RegisterCell RC = half(R2, 0).cat(half(R1, 0)).cat(half(R2, 1))
.cat(half(R1, 1));
return rr0(RC, Outputs);
}
case S2_shuffeb: {
RegisterCell RC = shuffle(rc(1), rc(2), 8, false);
return rr0(RC, Outputs);
}
case S2_shuffeh: {
RegisterCell RC = shuffle(rc(1), rc(2), 16, false);
return rr0(RC, Outputs);
}
case S2_shuffob: {
RegisterCell RC = shuffle(rc(1), rc(2), 8, true);
return rr0(RC, Outputs);
}
case S2_shuffoh: {
RegisterCell RC = shuffle(rc(1), rc(2), 16, true);
return rr0(RC, Outputs);
}
case C2_mask: {
uint16_t WR = W0;
uint16_t WP = 8; // XXX Pred size: getRegBitWidth(Reg[1]);
assert(WR == 64 && WP == 8);
RegisterCell R1 = rc(1);
RegisterCell RC(WR);
for (uint16_t i = 0; i < WP; ++i) {
const BT::BitValue &V = R1[i];
BT::BitValue F = (V.is(0) || V.is(1)) ? V : BT::BitValue::self();
RC.fill(i*8, i*8+8, F);
}
return rr0(RC, Outputs);
}
// Mux:
case C2_muxii:
case C2_muxir:
case C2_muxri:
case C2_mux: {
BT::BitValue PC0 = rc(1)[0];
RegisterCell R2 = cop(2, W0);
RegisterCell R3 = cop(3, W0);
if (PC0.is(0) || PC0.is(1))
return rr0(RegisterCell::ref(PC0 ? R2 : R3), Outputs);
R2.meet(R3, Reg[0].Reg);
return rr0(R2, Outputs);
}
case C2_vmux:
// TODO
break;
// Sign- and zero-extension:
case A2_sxtb:
return rr0(eSXT(rc(1), 8), Outputs);
case A2_sxth:
return rr0(eSXT(rc(1), 16), Outputs);
case A2_sxtw: {
uint16_t W1 = getRegBitWidth(Reg[1]);
assert(W0 == 64 && W1 == 32);
RegisterCell RC = eSXT(rc(1).cat(eIMM(0, W1)), W1);
return rr0(RC, Outputs);
}
case A2_zxtb:
return rr0(eZXT(rc(1), 8), Outputs);
case A2_zxth:
return rr0(eZXT(rc(1), 16), Outputs);
// Bit count:
case S2_cl0:
case S2_cl0p:
// Always produce a 32-bit result.
return rr0(eCLB(rc(1), 0/*bit*/, 32), Outputs);
case S2_cl1:
case S2_cl1p:
return rr0(eCLB(rc(1), 1/*bit*/, 32), Outputs);
case S2_clb:
case S2_clbp: {
uint16_t W1 = getRegBitWidth(Reg[1]);
RegisterCell R1 = rc(1);
BT::BitValue TV = R1[W1-1];
if (TV.is(0) || TV.is(1))
return rr0(eCLB(R1, TV, 32), Outputs);
break;
}
case S2_ct0:
case S2_ct0p:
return rr0(eCTB(rc(1), 0/*bit*/, 32), Outputs);
case S2_ct1:
case S2_ct1p:
return rr0(eCTB(rc(1), 1/*bit*/, 32), Outputs);
case S5_popcountp:
// TODO
break;
case C2_all8: {
RegisterCell P1 = rc(1);
bool Has0 = false, All1 = true;
for (uint16_t i = 0; i < 8/*XXX*/; ++i) {
if (!P1[i].is(1))
All1 = false;
if (!P1[i].is(0))
continue;
Has0 = true;
break;
}
if (!Has0 && !All1)
break;
RegisterCell RC(W0);
RC.fill(0, W0, (All1 ? BT::BitValue::One : BT::BitValue::Zero));
return rr0(RC, Outputs);
}
case C2_any8: {
RegisterCell P1 = rc(1);
bool Has1 = false, All0 = true;
for (uint16_t i = 0; i < 8/*XXX*/; ++i) {
if (!P1[i].is(0))
All0 = false;
if (!P1[i].is(1))
continue;
Has1 = true;
break;
}
if (!Has1 && !All0)
break;
RegisterCell RC(W0);
RC.fill(0, W0, (Has1 ? BT::BitValue::One : BT::BitValue::Zero));
return rr0(RC, Outputs);
}
case C2_and:
return rr0(eAND(rc(1), rc(2)), Outputs);
case C2_andn:
return rr0(eAND(rc(1), eNOT(rc(2))), Outputs);
case C2_not:
return rr0(eNOT(rc(1)), Outputs);
case C2_or:
return rr0(eORL(rc(1), rc(2)), Outputs);
case C2_orn:
return rr0(eORL(rc(1), eNOT(rc(2))), Outputs);
case C2_xor:
return rr0(eXOR(rc(1), rc(2)), Outputs);
case C4_and_and:
return rr0(eAND(rc(1), eAND(rc(2), rc(3))), Outputs);
case C4_and_andn:
return rr0(eAND(rc(1), eAND(rc(2), eNOT(rc(3)))), Outputs);
case C4_and_or:
return rr0(eAND(rc(1), eORL(rc(2), rc(3))), Outputs);
case C4_and_orn:
return rr0(eAND(rc(1), eORL(rc(2), eNOT(rc(3)))), Outputs);
case C4_or_and:
return rr0(eORL(rc(1), eAND(rc(2), rc(3))), Outputs);
case C4_or_andn:
return rr0(eORL(rc(1), eAND(rc(2), eNOT(rc(3)))), Outputs);
case C4_or_or:
return rr0(eORL(rc(1), eORL(rc(2), rc(3))), Outputs);
case C4_or_orn:
return rr0(eORL(rc(1), eORL(rc(2), eNOT(rc(3)))), Outputs);
case C2_bitsclr:
case C2_bitsclri:
case C2_bitsset:
case C4_nbitsclr:
case C4_nbitsclri:
case C4_nbitsset:
// TODO
break;
case S2_tstbit_i:
case S4_ntstbit_i: {
BT::BitValue V = rc(1)[im(2)];
if (V.is(0) || V.is(1)) {
// If instruction is S2_tstbit_i, test for 1, otherwise test for 0.
bool TV = (Opc == S2_tstbit_i);
BT::BitValue F = V.is(TV) ? BT::BitValue::One : BT::BitValue::Zero;
return rr0(RegisterCell(W0).fill(0, W0, F), Outputs);
}
break;
}
default:
return MachineEvaluator::evaluate(MI, Inputs, Outputs);
}
#undef im
#undef rc
#undef op
return false;
}
