//
// FIXME: select C, pow2, pow2 -> something smart
// FIXME: trunc(select X, Y, Z) -> select X, trunc(Y), trunc(Z)
-// FIXME: (select C, load A, load B) -> load (select C, A, B)
// FIXME: Dead stores -> nuke
-// FIXME: shr X, (and Y,31) -> shr X, Y
-// FIXME: TRUNC (LOAD) -> EXT_LOAD/LOAD(smaller)
+// FIXME: shr X, (and Y,31) -> shr X, Y (TRICKY!)
// FIXME: mul (x, const) -> shifts + adds
// FIXME: undef values
// FIXME: make truncate see through SIGN_EXTEND and AND
SDOperand visitSELECT(SDNode *N);
SDOperand visitSELECT_CC(SDNode *N);
SDOperand visitSETCC(SDNode *N);
+ SDOperand visitADD_PARTS(SDNode *N);
+ SDOperand visitSUB_PARTS(SDNode *N);
SDOperand visitSIGN_EXTEND(SDNode *N);
SDOperand visitZERO_EXTEND(SDNode *N);
SDOperand visitSIGN_EXTEND_INREG(SDNode *N);
SDOperand visitTRUNCATE(SDNode *N);
-
+
SDOperand visitFADD(SDNode *N);
SDOperand visitFSUB(SDNode *N);
SDOperand visitFMUL(SDNode *N);
SDOperand visitLOAD(SDNode *N);
SDOperand visitSTORE(SDNode *N);
+ bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS);
SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2);
SDOperand SimplifySelectCC(SDOperand N0, SDOperand N1, SDOperand N2,
SDOperand N3, ISD::CondCode CC);
SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
ISD::CondCode Cond, bool foldBooleans = true);
+
+ SDOperand BuildSDIV(SDNode *N);
+ SDOperand BuildUDIV(SDNode *N);
public:
DAGCombiner(SelectionDAG &D)
: DAG(D), TLI(D.getTargetLoweringInfo()), AfterLegalize(false) {}
};
}
+struct ms {
+ int64_t m; // magic number
+ int64_t s; // shift amount
+};
+
+struct mu {
+ uint64_t m; // magic number
+ int64_t a; // add indicator
+ int64_t s; // shift amount
+};
+
+/// magic - calculate the magic numbers required to codegen an integer sdiv as
+/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
+/// or -1.
+static ms magic32(int32_t d) {
+ int32_t p;
+ uint32_t ad, anc, delta, q1, r1, q2, r2, t;
+ const uint32_t two31 = 0x80000000U;
+ struct ms mag;
+
+ ad = abs(d);
+ t = two31 + ((uint32_t)d >> 31);
+ anc = t - 1 - t%ad; // absolute value of nc
+ p = 31; // initialize p
+ q1 = two31/anc; // initialize q1 = 2p/abs(nc)
+ r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
+ q2 = two31/ad; // initialize q2 = 2p/abs(d)
+ r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
+ do {
+ p = p + 1;
+ q1 = 2*q1; // update q1 = 2p/abs(nc)
+ r1 = 2*r1; // update r1 = rem(2p/abs(nc))
+ if (r1 >= anc) { // must be unsigned comparison
+ q1 = q1 + 1;
+ r1 = r1 - anc;
+ }
+ q2 = 2*q2; // update q2 = 2p/abs(d)
+ r2 = 2*r2; // update r2 = rem(2p/abs(d))
+ if (r2 >= ad) { // must be unsigned comparison
+ q2 = q2 + 1;
+ r2 = r2 - ad;
+ }
+ delta = ad - r2;
+ } while (q1 < delta || (q1 == delta && r1 == 0));
+
+ mag.m = (int32_t)(q2 + 1); // make sure to sign extend
+ if (d < 0) mag.m = -mag.m; // resulting magic number
+ mag.s = p - 32; // resulting shift
+ return mag;
+}
+
+/// magicu - calculate the magic numbers required to codegen an integer udiv as
+/// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
+static mu magicu32(uint32_t d) {
+ int32_t p;
+ uint32_t nc, delta, q1, r1, q2, r2;
+ struct mu magu;
+ magu.a = 0; // initialize "add" indicator
+ nc = - 1 - (-d)%d;
+ p = 31; // initialize p
+ q1 = 0x80000000/nc; // initialize q1 = 2p/nc
+ r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
+ q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
+ r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
+ do {
+ p = p + 1;
+ if (r1 >= nc - r1 ) {
+ q1 = 2*q1 + 1; // update q1
+ r1 = 2*r1 - nc; // update r1
+ }
+ else {
+ q1 = 2*q1; // update q1
+ r1 = 2*r1; // update r1
+ }
+ if (r2 + 1 >= d - r2) {
+ if (q2 >= 0x7FFFFFFF) magu.a = 1;
+ q2 = 2*q2 + 1; // update q2
+ r2 = 2*r2 + 1 - d; // update r2
+ }
+ else {
+ if (q2 >= 0x80000000) magu.a = 1;
+ q2 = 2*q2; // update q2
+ r2 = 2*r2 + 1; // update r2
+ }
+ delta = d - 1 - r2;
+ } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
+ magu.m = q2 + 1; // resulting magic number
+ magu.s = p - 32; // resulting shift
+ return magu;
+}
+
+/// magic - calculate the magic numbers required to codegen an integer sdiv as
+/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
+/// or -1.
+static ms magic64(int64_t d) {
+ int64_t p;
+ uint64_t ad, anc, delta, q1, r1, q2, r2, t;
+ const uint64_t two63 = 9223372036854775808ULL; // 2^63
+ struct ms mag;
+
+ ad = d >= 0 ? d : -d;
+ t = two63 + ((uint64_t)d >> 63);
+ anc = t - 1 - t%ad; // absolute value of nc
+ p = 63; // initialize p
+ q1 = two63/anc; // initialize q1 = 2p/abs(nc)
+ r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
+ q2 = two63/ad; // initialize q2 = 2p/abs(d)
+ r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
+ do {
+ p = p + 1;
+ q1 = 2*q1; // update q1 = 2p/abs(nc)
+ r1 = 2*r1; // update r1 = rem(2p/abs(nc))
+ if (r1 >= anc) { // must be unsigned comparison
+ q1 = q1 + 1;
+ r1 = r1 - anc;
+ }
+ q2 = 2*q2; // update q2 = 2p/abs(d)
+ r2 = 2*r2; // update r2 = rem(2p/abs(d))
+ if (r2 >= ad) { // must be unsigned comparison
+ q2 = q2 + 1;
+ r2 = r2 - ad;
+ }
+ delta = ad - r2;
+ } while (q1 < delta || (q1 == delta && r1 == 0));
+
+ mag.m = q2 + 1;
+ if (d < 0) mag.m = -mag.m; // resulting magic number
+ mag.s = p - 64; // resulting shift
+ return mag;
+}
+
+/// magicu - calculate the magic numbers required to codegen an integer udiv as
+/// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
+static mu magicu64(uint64_t d)
+{
+ int64_t p;
+ uint64_t nc, delta, q1, r1, q2, r2;
+ struct mu magu;
+ magu.a = 0; // initialize "add" indicator
+ nc = - 1 - (-d)%d;
+ p = 63; // initialize p
+ q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
+ r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
+ q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
+ r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
+ do {
+ p = p + 1;
+ if (r1 >= nc - r1 ) {
+ q1 = 2*q1 + 1; // update q1
+ r1 = 2*r1 - nc; // update r1
+ }
+ else {
+ q1 = 2*q1; // update q1
+ r1 = 2*r1; // update r1
+ }
+ if (r2 + 1 >= d - r2) {
+ if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
+ q2 = 2*q2 + 1; // update q2
+ r2 = 2*r2 + 1 - d; // update r2
+ }
+ else {
+ if (q2 >= 0x8000000000000000ull) magu.a = 1;
+ q2 = 2*q2; // update q2
+ r2 = 2*r2 + 1; // update r2
+ }
+ delta = d - 1 - r2;
+ } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
+ magu.m = q2 + 1; // resulting magic number
+ magu.s = p - 64; // resulting shift
+ return magu;
+}
+
/// MaskedValueIsZero - Return true if 'Op & Mask' is known to be zero. We use
/// this predicate to simplify operations downstream. Op and Mask are known to
/// be the same type.
case ISD::SELECT: return visitSELECT(N);
case ISD::SELECT_CC: return visitSELECT_CC(N);
case ISD::SETCC: return visitSETCC(N);
+ case ISD::ADD_PARTS: return visitADD_PARTS(N);
+ case ISD::SUB_PARTS: return visitSUB_PARTS(N);
case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
}
SDOperand DAGCombiner::visitTokenFactor(SDNode *N) {
+ std::vector<SDOperand> Ops;
+ bool Changed = false;
+
// If the token factor has two operands and one is the entry token, replace
// the token factor with the other operand.
if (N->getNumOperands() == 2) {
if (N->getOperand(1).getOpcode() == ISD::EntryToken)
return N->getOperand(0);
}
+
+ // fold (tokenfactor (tokenfactor)) -> tokenfactor
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
+ SDOperand Op = N->getOperand(i);
+ if (Op.getOpcode() == ISD::TokenFactor && Op.hasOneUse()) {
+ Changed = true;
+ for (unsigned j = 0, e = Op.getNumOperands(); j != e; ++j)
+ Ops.push_back(Op.getOperand(j));
+ } else {
+ Ops.push_back(Op);
+ }
+ }
+ if (Changed)
+ return DAG.getNode(ISD::TokenFactor, MVT::Other, Ops);
return SDOperand();
}
if (N0C && N1C)
return DAG.getConstant(N0C->getValue() + N1C->getValue(), VT);
// canonicalize constant to RHS
- if (N0C && !N1C) {
- std::swap(N0, N1);
- std::swap(N0C, N1C);
- }
+ if (N0C && !N1C)
+ return DAG.getNode(ISD::ADD, VT, N1, N0);
// fold (add x, 0) -> x
if (N1C && N1C->isNullValue())
return N0;
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
+ // fold (sub x, x) -> 0
+ if (N0 == N1)
+ return DAG.getConstant(0, N->getValueType(0));
+
// fold (sub c1, c2) -> c1-c2
if (N0C && N1C)
return DAG.getConstant(N0C->getValue() - N1C->getValue(),
return DAG.getConstant(N0C->getValue() * N1C->getValue(),
N->getValueType(0));
// canonicalize constant to RHS
- if (N0C && !N1C) {
- std::swap(N0, N1);
- std::swap(N0C, N1C);
- }
+ if (N0C && !N1C)
+ return DAG.getNode(ISD::MUL, VT, N1, N0);
// fold (mul x, 0) -> 0
if (N1C && N1C->isNullValue())
return N1;
// fold (mul x, -1) -> 0-x
if (N1C && N1C->isAllOnesValue())
- return DAG.getNode(ISD::SUB, N->getValueType(0),
- DAG.getConstant(0, N->getValueType(0)), N0);
+ return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
// fold (mul x, (1 << c)) -> x << c
if (N1C && isPowerOf2_64(N1C->getValue()))
return DAG.getNode(ISD::SHL, N->getValueType(0), N0,
if (N0C && N1C && !N1C->isNullValue())
return DAG.getConstant(N0C->getSignExtended() / N1C->getSignExtended(),
N->getValueType(0));
+ // fold (sdiv X, 1) -> X
+ if (N1C && N1C->getSignExtended() == 1LL)
+ return N0;
+ // fold (sdiv X, -1) -> 0-X
+ if (N1C && N1C->isAllOnesValue())
+ return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
// If we know the sign bits of both operands are zero, strength reduce to a
// udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
if (MaskedValueIsZero(N1, SignBit, TLI) &&
MaskedValueIsZero(N0, SignBit, TLI))
return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1);
+ // fold (sdiv X, pow2) -> (add (sra X, log(pow2)), (srl X, sizeof(X)-1))
+ if (N1C && N1C->getValue() && !TLI.isIntDivCheap() &&
+ (isPowerOf2_64(N1C->getSignExtended()) ||
+ isPowerOf2_64(-N1C->getSignExtended()))) {
+ // If dividing by powers of two is cheap, then don't perform the following
+ // fold.
+ if (TLI.isPow2DivCheap())
+ return SDOperand();
+ int64_t pow2 = N1C->getSignExtended();
+ int64_t abs2 = pow2 > 0 ? pow2 : -pow2;
+ SDOperand SRL = DAG.getNode(ISD::SRL, VT, N0,
+ DAG.getConstant(MVT::getSizeInBits(VT)-1,
+ TLI.getShiftAmountTy()));
+ WorkList.push_back(SRL.Val);
+ SDOperand SGN = DAG.getNode(ISD::ADD, VT, N0, SRL);
+ WorkList.push_back(SGN.Val);
+ SDOperand SRA = DAG.getNode(ISD::SRA, VT, SGN,
+ DAG.getConstant(Log2_64(abs2),
+ TLI.getShiftAmountTy()));
+ // If we're dividing by a positive value, we're done. Otherwise, we must
+ // negate the result.
+ if (pow2 > 0)
+ return SRA;
+ WorkList.push_back(SRA.Val);
+ return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), SRA);
+ }
+ // if integer divide is expensive and we satisfy the requirements, emit an
+ // alternate sequence.
+ if (N1C && (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) &&
+ !TLI.isIntDivCheap() &&
+ TLI.isOperationLegal(ISD::MULHS, VT) && TLI.isTypeLegal(VT)) {
+ return BuildSDIV(N);
+ }
return SDOperand();
}
SDOperand DAGCombiner::visitUDIV(SDNode *N) {
SDOperand N0 = N->getOperand(0);
SDOperand N1 = N->getOperand(1);
+ MVT::ValueType VT = N->getValueType(0);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
return DAG.getNode(ISD::SRL, N->getValueType(0), N0,
DAG.getConstant(Log2_64(N1C->getValue()),
TLI.getShiftAmountTy()));
+ // fold (udiv x, c) -> alternate
+ if (N1C && N1C->getValue() && TLI.isOperationLegal(ISD::MULHU, VT) &&
+ TLI.isTypeLegal(VT) && !TLI.isIntDivCheap())
+ return BuildUDIV(N);
return SDOperand();
}
if (N0C && N1C)
return DAG.getConstant(N0C->getValue() & N1C->getValue(), VT);
// canonicalize constant to RHS
- if (N0C && !N1C) {
- std::swap(N0, N1);
- std::swap(N0C, N1C);
- }
+ if (N0C && !N1C)
+ return DAG.getNode(ISD::AND, VT, N1, N0);
// fold (and x, -1) -> x
if (N1C && N1C->isAllOnesValue())
return N0;
WorkList.push_back(ANDNode.Val);
return DAG.getNode(N0.getOpcode(), VT, ANDNode, N0.getOperand(1));
}
+ // fold (and (sra)) -> (and (srl)) when possible.
+ if (N0.getOpcode() == ISD::SRA && N0.Val->hasOneUse())
+ if (ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
+ // If the RHS of the AND has zeros where the sign bits of the SRA will
+ // land, turn the SRA into an SRL.
+ if (MaskedValueIsZero(N1, (~0ULL << (OpSizeInBits-N01C->getValue())) &
+ (~0ULL>>(64-OpSizeInBits)), TLI)) {
+ WorkList.push_back(N);
+ CombineTo(N0.Val, DAG.getNode(ISD::SRL, VT, N0.getOperand(0),
+ N0.getOperand(1)));
+ return SDOperand();
+ }
+ }
+
+ // fold (zext_inreg (extload x)) -> (zextload x)
+ if (N0.getOpcode() == ISD::EXTLOAD) {
+ MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
+ // If we zero all the possible extended bits, then we can turn this into
+ // a zextload if we are running before legalize or the operation is legal.
+ if (MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT), TLI) &&
+ (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
+ SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
+ N0.getOperand(1), N0.getOperand(2),
+ EVT);
+ WorkList.push_back(N);
+ CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
+ return SDOperand();
+ }
+ }
+ // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
+ if (N0.getOpcode() == ISD::SEXTLOAD && N0.hasOneUse()) {
+ MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
+ // If we zero all the possible extended bits, then we can turn this into
+ // a zextload if we are running before legalize or the operation is legal.
+ if (MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT), TLI) &&
+ (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
+ SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
+ N0.getOperand(1), N0.getOperand(2),
+ EVT);
+ WorkList.push_back(N);
+ CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
+ return SDOperand();
+ }
+ }
return SDOperand();
}
return DAG.getConstant(N0C->getValue() | N1C->getValue(),
N->getValueType(0));
// canonicalize constant to RHS
- if (N0C && !N1C) {
- std::swap(N0, N1);
- std::swap(N0C, N1C);
- }
+ if (N0C && !N1C)
+ return DAG.getNode(ISD::OR, VT, N1, N0);
// fold (or x, 0) -> x
if (N1C && N1C->isNullValue())
return N0;
if (N0C && N1C)
return DAG.getConstant(N0C->getValue() ^ N1C->getValue(), VT);
// canonicalize constant to RHS
- if (N0C && !N1C) {
- std::swap(N0, N1);
- std::swap(N0C, N1C);
- }
+ if (N0C && !N1C)
+ return DAG.getNode(ISD::XOR, VT, N1, N0);
// fold (xor x, 0) -> x
if (N1C && N1C->isNullValue())
return N0;
// fold X ? Y : X --> X ? Y : 0 --> X & Y
if (MVT::i1 == VT && N0 == N2)
return DAG.getNode(ISD::AND, VT, N0, N1);
+
+ // If we can fold this based on the true/false value, do so.
+ if (SimplifySelectOps(N, N1, N2))
+ return SDOperand();
+
// fold selects based on a setcc into other things, such as min/max/abs
if (N0.getOpcode() == ISD::SETCC)
return SimplifySelect(N0, N1, N2);
// fold select_cc lhs, rhs, x, x, cc -> x
if (N2 == N3)
return N2;
+
+ // If we can fold this based on the true/false value, do so.
+ if (SimplifySelectOps(N, N2, N3))
+ return SDOperand();
+
// fold select_cc into other things, such as min/max/abs
return SimplifySelectCC(N0, N1, N2, N3, CC);
}
cast<CondCodeSDNode>(N->getOperand(2))->get());
}
+SDOperand DAGCombiner::visitADD_PARTS(SDNode *N) {
+ SDOperand LHSLo = N->getOperand(0);
+ SDOperand RHSLo = N->getOperand(2);
+ MVT::ValueType VT = LHSLo.getValueType();
+
+ // fold (a_Hi, 0) + (b_Hi, b_Lo) -> (b_Hi + a_Hi, b_Lo)
+ if (MaskedValueIsZero(LHSLo, (1ULL << MVT::getSizeInBits(VT))-1, TLI)) {
+ SDOperand Hi = DAG.getNode(ISD::ADD, VT, N->getOperand(1),
+ N->getOperand(3));
+ WorkList.push_back(Hi.Val);
+ CombineTo(N, RHSLo, Hi);
+ return SDOperand();
+ }
+ // fold (a_Hi, a_Lo) + (b_Hi, 0) -> (a_Hi + b_Hi, a_Lo)
+ if (MaskedValueIsZero(RHSLo, (1ULL << MVT::getSizeInBits(VT))-1, TLI)) {
+ SDOperand Hi = DAG.getNode(ISD::ADD, VT, N->getOperand(1),
+ N->getOperand(3));
+ WorkList.push_back(Hi.Val);
+ CombineTo(N, LHSLo, Hi);
+ return SDOperand();
+ }
+ return SDOperand();
+}
+
+SDOperand DAGCombiner::visitSUB_PARTS(SDNode *N) {
+ SDOperand LHSLo = N->getOperand(0);
+ SDOperand RHSLo = N->getOperand(2);
+ MVT::ValueType VT = LHSLo.getValueType();
+
+ // fold (a_Hi, a_Lo) - (b_Hi, 0) -> (a_Hi - b_Hi, a_Lo)
+ if (MaskedValueIsZero(RHSLo, (1ULL << MVT::getSizeInBits(VT))-1, TLI)) {
+ SDOperand Hi = DAG.getNode(ISD::SUB, VT, N->getOperand(1),
+ N->getOperand(3));
+ WorkList.push_back(Hi.Val);
+ CombineTo(N, LHSLo, Hi);
+ return SDOperand();
+ }
+ return SDOperand();
+}
+
SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
SDOperand N0 = N->getOperand(0);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
if (N0.getOpcode() == ISD::SEXTLOAD && VT == N0.getValueType())
return N0;
// fold (sext (load x)) -> (sextload x)
- if (N0.getOpcode() == ISD::LOAD && N0.Val->hasNUsesOfValue(1, 0)) {
+ if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) {
SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
N0.getOperand(1), N0.getOperand(2),
N0.getValueType());
- CombineTo(N0.Val, ExtLoad, ExtLoad.getOperand(0));
WorkList.push_back(N);
+ CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
+ ExtLoad.getValue(1));
return SDOperand();
}
return SDOperand();
}
// fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt1
if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
- cast<VTSDNode>(N0.getOperand(1))->getVT() < EVT) {
+ cast<VTSDNode>(N0.getOperand(1))->getVT() <= EVT) {
return N0;
}
// fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
return DAG.getNode(ISD::SRA, N0.getValueType(), N0.getOperand(0),
N0.getOperand(1));
}
+ // fold (sext_inreg (extload x)) -> (sextload x)
+ if (N0.getOpcode() == ISD::EXTLOAD &&
+ EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() &&
+ (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) {
+ SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
+ N0.getOperand(1), N0.getOperand(2),
+ EVT);
+ WorkList.push_back(N);
+ CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
+ return SDOperand();
+ }
+ // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
+ if (N0.getOpcode() == ISD::ZEXTLOAD && N0.hasOneUse() &&
+ EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() &&
+ (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) {
+ SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
+ N0.getOperand(1), N0.getOperand(2),
+ EVT);
+ WorkList.push_back(N);
+ CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
+ return SDOperand();
+ }
return SDOperand();
}
return N0.getOperand(0);
}
// fold (truncate (load x)) -> (smaller load x)
- if (N0.getOpcode() == ISD::LOAD && N0.Val->hasNUsesOfValue(1, 0)) {
+ if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) {
assert(MVT::getSizeInBits(N0.getValueType()) > MVT::getSizeInBits(VT) &&
"Cannot truncate to larger type!");
MVT::ValueType PtrType = N0.getOperand(1).getValueType();
DAG.getConstant(PtrOff, PtrType));
WorkList.push_back(NewPtr.Val);
SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), NewPtr,N0.getOperand(2));
- CombineTo(N0.Val, Load, Load.getOperand(0));
WorkList.push_back(N);
+ CombineTo(N0.Val, Load, Load.getValue(1));
return SDOperand();
}
return SDOperand();
SDOperand DAGCombiner::visitFADD(SDNode *N) {
SDOperand N0 = N->getOperand(0);
SDOperand N1 = N->getOperand(1);
+ ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
+ ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
MVT::ValueType VT = N->getValueType(0);
-
- if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0))
- if (ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1)) {
- // fold floating point (fadd c1, c2)
- return DAG.getConstantFP(N0CFP->getValue() + N1CFP->getValue(),
- N->getValueType(0));
- }
+
+ // fold (fadd c1, c2) -> c1+c2
+ if (N0CFP && N1CFP)
+ return DAG.getConstantFP(N0CFP->getValue() + N1CFP->getValue(), VT);
+ // canonicalize constant to RHS
+ if (N0CFP && !N1CFP)
+ return DAG.getNode(ISD::FADD, VT, N1, N0);
// fold (A + (-B)) -> A-B
if (N1.getOpcode() == ISD::FNEG)
return DAG.getNode(ISD::FSUB, VT, N0, N1.getOperand(0));
-
// fold ((-A) + B) -> B-A
if (N0.getOpcode() == ISD::FNEG)
return DAG.getNode(ISD::FSUB, VT, N1, N0.getOperand(0));
-
return SDOperand();
}
SDOperand DAGCombiner::visitFSUB(SDNode *N) {
SDOperand N0 = N->getOperand(0);
SDOperand N1 = N->getOperand(1);
+ ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
+ ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
MVT::ValueType VT = N->getValueType(0);
-
- if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0))
- if (ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1)) {
- // fold floating point (fsub c1, c2)
- return DAG.getConstantFP(N0CFP->getValue() - N1CFP->getValue(),
- N->getValueType(0));
- }
+
+ // fold (fsub c1, c2) -> c1-c2
+ if (N0CFP && N1CFP)
+ return DAG.getConstantFP(N0CFP->getValue() - N1CFP->getValue(), VT);
// fold (A-(-B)) -> A+B
if (N1.getOpcode() == ISD::FNEG)
return DAG.getNode(ISD::FADD, N0.getValueType(), N0, N1.getOperand(0));
-
return SDOperand();
}
SDOperand DAGCombiner::visitFMUL(SDNode *N) {
SDOperand N0 = N->getOperand(0);
SDOperand N1 = N->getOperand(1);
+ ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
+ ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
MVT::ValueType VT = N->getValueType(0);
- if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0))
- if (ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1)) {
- // fold floating point (fmul c1, c2)
- return DAG.getConstantFP(N0CFP->getValue() * N1CFP->getValue(),
- N->getValueType(0));
- }
+ // fold (fmul c1, c2) -> c1*c2
+ if (N0CFP && N1CFP)
+ return DAG.getConstantFP(N0CFP->getValue() * N1CFP->getValue(), VT);
+ // canonicalize constant to RHS
+ if (N0CFP && !N1CFP)
+ return DAG.getNode(ISD::FMUL, VT, N1, N0);
+ // fold (fmul X, 2.0) -> (fadd X, X)
+ if (N1CFP && N1CFP->isExactlyValue(+2.0))
+ return DAG.getNode(ISD::FADD, VT, N0, N0);
return SDOperand();
}
if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0))
if (ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1)) {
// fold floating point (fdiv c1, c2)
- return DAG.getConstantFP(N0CFP->getValue() / N1CFP->getValue(),
- N->getValueType(0));
+ return DAG.getConstantFP(N0CFP->getValue() / N1CFP->getValue(), VT);
}
return SDOperand();
}
if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0))
if (ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1)) {
// fold floating point (frem c1, c2) -> fmod(c1, c2)
- return DAG.getConstantFP(fmod(N0CFP->getValue(),N1CFP->getValue()),
- N->getValueType(0));
+ return DAG.getConstantFP(fmod(N0CFP->getValue(),N1CFP->getValue()), VT);
}
return SDOperand();
}
return SDOperand();
}
+/// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
+/// are the two values being selected between, see if we can simplify the
+/// select.
+///
+bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDOperand LHS,
+ SDOperand RHS) {
+
+ // If this is a select from two identical things, try to pull the operation
+ // through the select.
+ if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){
+#if 0
+ std::cerr << "SELECT: ["; LHS.Val->dump();
+ std::cerr << "] ["; RHS.Val->dump();
+ std::cerr << "]\n";
+#endif
+
+ // If this is a load and the token chain is identical, replace the select
+ // of two loads with a load through a select of the address to load from.
+ // This triggers in things like "select bool X, 10.0, 123.0" after the FP
+ // constants have been dropped into the constant pool.
+ if ((LHS.getOpcode() == ISD::LOAD ||
+ LHS.getOpcode() == ISD::EXTLOAD ||
+ LHS.getOpcode() == ISD::ZEXTLOAD ||
+ LHS.getOpcode() == ISD::SEXTLOAD) &&
+ // Token chains must be identical.
+ LHS.getOperand(0) == RHS.getOperand(0) &&
+ // If this is an EXTLOAD, the VT's must match.
+ (LHS.getOpcode() == ISD::LOAD ||
+ LHS.getOperand(3) == RHS.getOperand(3))) {
+ // FIXME: this conflates two src values, discarding one. This is not
+ // the right thing to do, but nothing uses srcvalues now. When they do,
+ // turn SrcValue into a list of locations.
+ SDOperand Addr;
+ if (TheSelect->getOpcode() == ISD::SELECT)
+ Addr = DAG.getNode(ISD::SELECT, LHS.getOperand(1).getValueType(),
+ TheSelect->getOperand(0), LHS.getOperand(1),
+ RHS.getOperand(1));
+ else
+ Addr = DAG.getNode(ISD::SELECT_CC, LHS.getOperand(1).getValueType(),
+ TheSelect->getOperand(0),
+ TheSelect->getOperand(1),
+ LHS.getOperand(1), RHS.getOperand(1),
+ TheSelect->getOperand(4));
+
+ SDOperand Load;
+ if (LHS.getOpcode() == ISD::LOAD)
+ Load = DAG.getLoad(TheSelect->getValueType(0), LHS.getOperand(0),
+ Addr, LHS.getOperand(2));
+ else
+ Load = DAG.getExtLoad(LHS.getOpcode(), TheSelect->getValueType(0),
+ LHS.getOperand(0), Addr, LHS.getOperand(2),
+ cast<VTSDNode>(LHS.getOperand(3))->getVT());
+ // Users of the select now use the result of the load.
+ CombineTo(TheSelect, Load);
+
+ // Users of the old loads now use the new load's chain. We know the
+ // old-load value is dead now.
+ CombineTo(LHS.Val, Load.getValue(0), Load.getValue(1));
+ CombineTo(RHS.Val, Load.getValue(0), Load.getValue(1));
+ return true;
+ }
+ }
+
+ return false;
+}
+
SDOperand DAGCombiner::SimplifySelectCC(SDOperand N0, SDOperand N1,
SDOperand N2, SDOperand N3,
ISD::CondCode CC) {
return SDOperand();
}
+/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
+/// return a DAG expression to select that will generate the same value by
+/// multiplying by a magic number. See:
+/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
+SDOperand DAGCombiner::BuildSDIV(SDNode *N) {
+ MVT::ValueType VT = N->getValueType(0);
+ assert((VT == MVT::i32 || VT == MVT::i64) &&
+ "BuildSDIV only operates on i32 or i64!");
+
+ int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended();
+ ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
+
+ // Multiply the numerator (operand 0) by the magic value
+ SDOperand Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
+ DAG.getConstant(magics.m, VT));
+ // If d > 0 and m < 0, add the numerator
+ if (d > 0 && magics.m < 0) {
+ Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
+ WorkList.push_back(Q.Val);
+ }
+ // If d < 0 and m > 0, subtract the numerator.
+ if (d < 0 && magics.m > 0) {
+ Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
+ WorkList.push_back(Q.Val);
+ }
+ // Shift right algebraic if shift value is nonzero
+ if (magics.s > 0) {
+ Q = DAG.getNode(ISD::SRA, VT, Q,
+ DAG.getConstant(magics.s, TLI.getShiftAmountTy()));
+ WorkList.push_back(Q.Val);
+ }
+ // Extract the sign bit and add it to the quotient
+ SDOperand T =
+ DAG.getNode(ISD::SRL, MVT::i32, Q,
+ DAG.getConstant(MVT::getSizeInBits(VT)-1,
+ TLI.getShiftAmountTy()));
+ WorkList.push_back(T.Val);
+ return DAG.getNode(ISD::ADD, VT, Q, T);
+}
+
+/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
+/// return a DAG expression to select that will generate the same value by
+/// multiplying by a magic number. See:
+/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
+SDOperand DAGCombiner::BuildUDIV(SDNode *N) {
+ MVT::ValueType VT = N->getValueType(0);
+ assert((VT == MVT::i32 || VT == MVT::i64) &&
+ "BuildUDIV only operates on i32 or i64!");
+
+ uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
+ mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
+
+ // Multiply the numerator (operand 0) by the magic value
+ SDOperand Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
+ DAG.getConstant(magics.m, VT));
+ WorkList.push_back(Q.Val);
+
+ if (magics.a == 0) {
+ return DAG.getNode(ISD::SRL, VT, Q,
+ DAG.getConstant(magics.s, TLI.getShiftAmountTy()));
+ } else {
+ SDOperand NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
+ WorkList.push_back(NPQ.Val);
+ NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
+ DAG.getConstant(1, TLI.getShiftAmountTy()));
+ WorkList.push_back(NPQ.Val);
+ NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
+ WorkList.push_back(NPQ.Val);
+ return DAG.getNode(ISD::SRL, VT, NPQ,
+ DAG.getConstant(magics.s-1, TLI.getShiftAmountTy()));
+ }
+}
+
// SelectionDAG::Combine - This is the entry point for the file.
//
void SelectionDAG::Combine(bool RunningAfterLegalize) {