// ZERO_EXTEND/SIGN_EXTEND by converting them to an ANY_EXTEND node which
// we don't have yet.
//
+// FIXME: select C, pow2, pow2 -> something smart
+// FIXME: trunc(select X, Y, Z) -> select X, trunc(Y), trunc(Z)
+// FIXME: Dead stores -> nuke
+// FIXME: shr X, (and Y,31) -> shr X, Y (TRICKY!)
// FIXME: mul (x, const) -> shifts + adds
// FIXME: undef values
-// FIXME: zero extend when top bits are 0 -> drop it ?
// FIXME: make truncate see through SIGN_EXTEND and AND
-// FIXME: sext_in_reg(setcc) on targets that return zero or one, and where
-// EVT != MVT::i1 can drop the sext.
// FIXME: (sra (sra x, c1), c2) -> (sra x, c1+c2)
// FIXME: verify that getNode can't return extends with an operand whose type
// is >= to that of the extend.
// FIXME: divide by zero is currently left unfolded. do we want to turn this
// into an undef?
+// FIXME: select ne (select cc, 1, 0), 0, true, false -> select cc, true, false
+// FIXME: reassociate (X+C)+Y into (X+Y)+C if the inner expression has one use
//
//===----------------------------------------------------------------------===//
WorkList.end());
}
+ SDOperand CombineTo(SDNode *N, const std::vector<SDOperand> &To) {
+ ++NodesCombined;
+ DEBUG(std::cerr << "\nReplacing "; N->dump();
+ std::cerr << "\nWith: "; To[0].Val->dump();
+ std::cerr << " and " << To.size()-1 << " other values\n");
+ std::vector<SDNode*> NowDead;
+ DAG.ReplaceAllUsesWith(N, To, &NowDead);
+
+ // Push the new nodes and any users onto the worklist
+ for (unsigned i = 0, e = To.size(); i != e; ++i) {
+ WorkList.push_back(To[i].Val);
+ AddUsersToWorkList(To[i].Val);
+ }
+
+ // Nodes can end up on the worklist more than once. Make sure we do
+ // not process a node that has been replaced.
+ removeFromWorkList(N);
+ for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
+ removeFromWorkList(NowDead[i]);
+
+ // Finally, since the node is now dead, remove it from the graph.
+ DAG.DeleteNode(N);
+ return SDOperand(N, 0);
+ }
+
+ SDOperand CombineTo(SDNode *N, SDOperand Res) {
+ std::vector<SDOperand> To;
+ To.push_back(Res);
+ return CombineTo(N, To);
+ }
+
+ SDOperand CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) {
+ std::vector<SDOperand> To;
+ To.push_back(Res0);
+ To.push_back(Res1);
+ return CombineTo(N, To);
+ }
+
/// visit - call the node-specific routine that knows how to fold each
/// particular type of node.
SDOperand visit(SDNode *N);
// node types. The semantics are as follows:
// Return Value:
// SDOperand.Val == 0 - No change was made
+ // SDOperand.Val == N - N was replaced, is dead, and is already handled.
// otherwise - N should be replaced by the returned Operand.
//
SDOperand visitTokenFactor(SDNode *N);
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 visitFDIV(SDNode *N);
+ SDOperand visitFREM(SDNode *N);
SDOperand visitSINT_TO_FP(SDNode *N);
SDOperand visitUINT_TO_FP(SDNode *N);
SDOperand visitFP_TO_SINT(SDNode *N);
SDOperand visitFNEG(SDNode *N);
SDOperand visitFABS(SDNode *N);
SDOperand visitBRCOND(SDNode *N);
- // brcondtwoway
- // br_cc
- // brtwoway_cc
+ SDOperand visitBRCONDTWOWAY(SDNode *N);
+ SDOperand visitBR_CC(SDNode *N);
+ SDOperand visitBRTWOWAY_CC(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);
+ 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) {}
};
}
-/// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
-/// this predicate to simplify operations downstream. V and Mask are known to
+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.
static bool MaskedValueIsZero(const SDOperand &Op, uint64_t Mask,
const TargetLowering &TLI) {
case ISD::Constant:
return (cast<ConstantSDNode>(Op)->getValue() & Mask) == 0;
case ISD::SETCC:
- // FIXME: teach this about non ZeroOrOne values, such as 0 or -1
return ((Mask & 1) == 0) &&
TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult;
case ISD::ZEXTLOAD:
SrcBits = MVT::getSizeInBits(cast<VTSDNode>(Op.getOperand(1))->getVT());
return (Mask & ((1ULL << SrcBits)-1)) == 0; // Returning only the zext bits.
case ISD::AND:
+ // If either of the operands has zero bits, the result will too.
+ if (MaskedValueIsZero(Op.getOperand(1), Mask, TLI) ||
+ MaskedValueIsZero(Op.getOperand(0), Mask, TLI))
+ return true;
// (X & C1) & C2 == 0 iff C1 & C2 == 0.
if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
return MaskedValueIsZero(Op.getOperand(0),AndRHS->getValue() & Mask, TLI);
- // FALL THROUGH
+ return false;
case ISD::OR:
case ISD::XOR:
return MaskedValueIsZero(Op.getOperand(0), Mask, TLI) &&
return MaskedValueIsZero(Op.getOperand(0), NewVal, TLI);
}
return false;
+ case ISD::ADD:
+ // (add X, Y) & C == 0 iff (X&C)|(Y&C) == 0 and all bits are low bits.
+ if ((Mask&(Mask+1)) == 0) { // All low bits
+ if (MaskedValueIsZero(Op.getOperand(0), Mask, TLI) &&
+ MaskedValueIsZero(Op.getOperand(1), Mask, TLI))
+ return true;
+ }
+ break;
+ case ISD::SUB:
+ if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) {
+ // We know that the top bits of C-X are clear if X contains less bits
+ // than C (i.e. no wrap-around can happen). For example, 20-X is
+ // positive if we can prove that X is >= 0 and < 16.
+ unsigned Bits = MVT::getSizeInBits(CLHS->getValueType(0));
+ if ((CLHS->getValue() & (1 << (Bits-1))) == 0) { // sign bit clear
+ unsigned NLZ = CountLeadingZeros_64(CLHS->getValue()+1);
+ uint64_t MaskV = (1ULL << (63-NLZ))-1;
+ if (MaskedValueIsZero(Op.getOperand(1), ~MaskV, TLI)) {
+ // High bits are clear this value is known to be >= C.
+ unsigned NLZ2 = CountLeadingZeros_64(CLHS->getValue());
+ if ((Mask & ((1ULL << (64-NLZ2))-1)) == 0)
+ return true;
+ }
+ }
+ }
+ break;
case ISD::CTTZ:
case ISD::CTLZ:
case ISD::CTPOP:
// Bit counting instructions can not set the high bits of the result
// register. The max number of bits sets depends on the input.
return (Mask & (MVT::getSizeInBits(Op.getValueType())*2-1)) == 0;
-
- // TODO we could handle some SRA cases here.
default: break;
}
return false;
// Add all the dag nodes to the worklist.
WorkList.insert(WorkList.end(), DAG.allnodes_begin(), DAG.allnodes_end());
+ // Create a dummy node (which is not added to allnodes), that adds a reference
+ // to the root node, preventing it from being deleted, and tracking any
+ // changes of the root.
+ HandleSDNode Dummy(DAG.getRoot());
+
// while the worklist isn't empty, inspect the node on the end of it and
// try and combine it.
while (!WorkList.empty()) {
WorkList.pop_back();
// If N has no uses, it is dead. Make sure to revisit all N's operands once
- // N is deleted from the DAG, since they too may now be dead.
- // FIXME: is there a better way to keep from deleting the dag root because
- // we think it has no uses? This works for now...
- if (N->use_empty() && N != DAG.getRoot().Val) {
+ // N is deleted from the DAG, since they too may now be dead or may have a
+ // reduced number of uses, allowing other xforms.
+ if (N->use_empty() && N != &Dummy) {
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
WorkList.push_back(N->getOperand(i).Val);
- DAG.DeleteNode(N);
removeFromWorkList(N);
+ DAG.DeleteNode(N);
continue;
}
DEBUG(std::cerr << "\nReplacing "; N->dump();
std::cerr << "\nWith: "; RV.Val->dump();
std::cerr << '\n');
- DAG.ReplaceAllUsesWith(N, std::vector<SDOperand>(1, RV));
+ std::vector<SDNode*> NowDead;
+ DAG.ReplaceAllUsesWith(N, std::vector<SDOperand>(1, RV), &NowDead);
// Push the new node and any users onto the worklist
WorkList.push_back(RV.Val);
// Nodes can end up on the worklist more than once. Make sure we do
// not process a node that has been replaced.
removeFromWorkList(N);
+ for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
+ removeFromWorkList(NowDead[i]);
+
+ // Finally, since the node is now dead, remove it from the graph.
+ DAG.DeleteNode(N);
}
}
}
+
+ // If the root changed (e.g. it was a dead load, update the root).
+ DAG.setRoot(Dummy.getValue());
}
SDOperand DAGCombiner::visit(SDNode *N) {
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);
case ISD::TRUNCATE: return visitTRUNCATE(N);
+ case ISD::FADD: return visitFADD(N);
+ case ISD::FSUB: return visitFSUB(N);
+ case ISD::FMUL: return visitFMUL(N);
+ case ISD::FDIV: return visitFDIV(N);
+ case ISD::FREM: return visitFREM(N);
case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
case ISD::FP_EXTEND: return visitFP_EXTEND(N);
case ISD::FNEG: return visitFNEG(N);
case ISD::FABS: return visitFABS(N);
+ case ISD::BRCOND: return visitBRCOND(N);
+ case ISD::BRCONDTWOWAY: return visitBRCONDTWOWAY(N);
+ case ISD::BR_CC: return visitBR_CC(N);
+ case ISD::BRTWOWAY_CC: return visitBRTWOWAY_CC(N);
+ case ISD::LOAD: return visitLOAD(N);
+ case ISD::STORE: return visitSTORE(N);
}
return SDOperand();
}
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();
}
SDOperand N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
- ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
MVT::ValueType VT = N0.getValueType();
// fold (add c1, c2) -> c1+c2
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;
- // fold floating point (add c1, c2) -> c1+c2
- if (N0CFP && N1CFP)
- return DAG.getConstantFP(N0CFP->getValue() + N1CFP->getValue(), VT);
// fold (add (add x, c1), c2) -> (add x, c1+c2)
if (N1C && N0.getOpcode() == ISD::ADD) {
ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
return DAG.getNode(ISD::ADD, VT, N0.getOperand(0),
DAG.getConstant(N1C->getValue()+N01C->getValue(), VT));
}
- // fold (A + (-B)) -> A-B
- if (N1.getOpcode() == ISD::FNEG)
- return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(0));
- // fold ((-A) + B) -> B-A
- if (N0.getOpcode() == ISD::FNEG)
- return DAG.getNode(ISD::SUB, VT, N1, N0.getOperand(0));
// fold ((0-A) + B) -> B-A
if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(1));
- // fold (A+(B-A)) -> B for non-fp types
- if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1) &&
- !MVT::isFloatingPoint(N1.getValueType()))
+ // fold (A+(B-A)) -> B
+ if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
return N1.getOperand(0);
return SDOperand();
}
SDOperand N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
- ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0.Val);
- ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(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(),
N->getValueType(0));
- // fold (sub x, 0) -> x
- if (N1C && N1C->isNullValue())
- return N0;
- // fold floating point (sub c1, c2) -> c1-c2
- if (N0CFP && N1CFP)
- return DAG.getConstantFP(N0CFP->getValue() - N1CFP->getValue(),
- N->getValueType(0));
+ // fold (sub x, c) -> (add x, -c)
+ if (N1C)
+ return DAG.getNode(ISD::ADD, N0.getValueType(), N0,
+ DAG.getConstant(-N1C->getValue(), N0.getValueType()));
+
// fold (A+B)-A -> B
- if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
- !MVT::isFloatingPoint(N1.getValueType()))
+ if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
return N0.getOperand(1);
// fold (A+B)-B -> A
- if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
- !MVT::isFloatingPoint(N1.getValueType()))
+ if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
return N0.getOperand(0);
- // fold (A-(-B)) -> A+B
- if (N1.getOpcode() == ISD::FNEG)
- return DAG.getNode(ISD::ADD, N0.getValueType(), N0, N1.getOperand(0));
return SDOperand();
}
SDOperand N1 = N->getOperand(1);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
- ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
MVT::ValueType VT = N0.getValueType();
// fold (mul c1, c2) -> c1*c2
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,
return DAG.getNode(ISD::MUL, VT, N0.getOperand(0),
DAG.getConstant(N1C->getValue()*N01C->getValue(), VT));
}
- // fold floating point (mul c1, c2) -> c1*c2
- if (N0CFP && N1CFP)
- return DAG.getConstantFP(N0CFP->getValue() * N1CFP->getValue(),
- N->getValueType(0));
return SDOperand();
}
SDOperand DAGCombiner::visitSDIV(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);
- ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0.Val);
- ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1.Val);
// fold (sdiv c1, c2) -> c1/c2
if (N0C && N1C && !N1C->isNullValue())
return DAG.getConstant(N0C->getSignExtended() / N1C->getSignExtended(),
N->getValueType(0));
- // fold floating point (sdiv c1, c2) -> c1/c2
- if (N0CFP && N1CFP)
- return DAG.getConstantFP(N0CFP->getValue() / N1CFP->getValue(),
- 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();
}
SDOperand DAGCombiner::visitSREM(SDNode *N) {
SDOperand N0 = N->getOperand(0);
SDOperand N1 = N->getOperand(1);
+ MVT::ValueType VT = N->getValueType(0);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
- ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
- ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
// fold (srem c1, c2) -> c1%c2
if (N0C && N1C && !N1C->isNullValue())
return DAG.getConstant(N0C->getSignExtended() % N1C->getSignExtended(),
N->getValueType(0));
- // fold floating point (srem c1, c2) -> fmod(c1, c2)
- if (N0CFP && N1CFP)
- return DAG.getConstantFP(fmod(N0CFP->getValue(),N1CFP->getValue()),
- N->getValueType(0));
+ // If we know the sign bits of both operands are zero, strength reduce to a
+ // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
+ uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
+ if (MaskedValueIsZero(N1, SignBit, TLI) &&
+ MaskedValueIsZero(N0, SignBit, TLI))
+ return DAG.getNode(ISD::UREM, N1.getValueType(), N0, N1);
return SDOperand();
}
if (N0C && N1C && !N1C->isNullValue())
return DAG.getConstant(N0C->getValue() % N1C->getValue(),
N->getValueType(0));
- // FIXME: c2 power of 2 -> mask?
+ // fold (urem x, pow2) -> (and x, pow2-1)
+ if (N1C && !N1C->isNullValue() && isPowerOf2_64(N1C->getValue()))
+ return DAG.getNode(ISD::AND, N0.getValueType(), N0,
+ DAG.getConstant(N1C->getValue()-1, N1.getValueType()));
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;
return DAG.getNode(ISD::AND, VT, N0.getOperand(0), N1);
}
// fold (and (or x, 0xFFFF), 0xFF) -> 0xFF
- if (N0.getOpcode() == ISD::OR)
+ if (N0.getOpcode() == ISD::OR && N1C)
if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
if ((ORI->getValue() & N1C->getValue()) == N1C->getValue())
return N1;
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;
if (N1C && N1C->isNullValue())
return N0;
// If the sign bit is known to be zero, switch this to a SRL.
- if (N1C && MaskedValueIsZero(N0, (1ULL << (OpSizeInBits-1)), TLI))
+ if (MaskedValueIsZero(N0, (1ULL << (OpSizeInBits-1)), TLI))
return DAG.getNode(ISD::SRL, VT, N0, N1);
return SDOperand();
}
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
MVT::ValueType VT = N->getValueType(0);
-
+
// fold select C, X, X -> X
if (N1 == N2)
return N1;
if (N0C && N0C->isNullValue())
return N2;
// fold select C, 1, X -> C | X
- if (MVT::i1 == VT && N1C && !N1C->isNullValue())
+ if (MVT::i1 == VT && N1C && N1C->getValue() == 1)
return DAG.getNode(ISD::OR, VT, N0, N2);
// fold select C, 0, X -> ~C & X
// FIXME: this should check for C type == X type, not i1?
return DAG.getNode(ISD::AND, VT, XORNode, N2);
}
// fold select C, X, 1 -> ~C | X
- if (MVT::i1 == VT && N2C && !N2C->isNullValue()) {
+ if (MVT::i1 == VT && N2C && N2C->getValue() == 1) {
SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
WorkList.push_back(XORNode.Val);
return DAG.getNode(ISD::OR, VT, XORNode, N1);
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);
return SDOperand();
}
SDOperand DAGCombiner::visitSELECT_CC(SDNode *N) {
- return SDOperand();
+ SDOperand N0 = N->getOperand(0);
+ SDOperand N1 = N->getOperand(1);
+ SDOperand N2 = N->getOperand(2);
+ SDOperand N3 = N->getOperand(3);
+ SDOperand N4 = N->getOperand(4);
+ ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
+ ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
+ ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
+ ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
+
+ // Determine if the condition we're dealing with is constant
+ SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
+ ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
+
+ // 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);
}
SDOperand DAGCombiner::visitSETCC(SDNode *N) {
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);
// fold (sext (sext x)) -> (sext x)
if (N0.getOpcode() == ISD::SIGN_EXTEND)
return DAG.getNode(ISD::SIGN_EXTEND, VT, N0.getOperand(0));
+ // fold (sext (sextload x)) -> (sextload x)
+ if (N0.getOpcode() == ISD::SEXTLOAD && VT == N0.getValueType())
+ return N0;
+ // fold (sext (load x)) -> (sextload x)
+ 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());
+ WorkList.push_back(N);
+ CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
+ ExtLoad.getValue(1));
+ return SDOperand();
+ }
return SDOperand();
}
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
MVT::ValueType VT = N->getValueType(0);
MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT();
+ unsigned EVTBits = MVT::getSizeInBits(EVT);
// fold (sext_in_reg c1) -> c1
if (N0C) {
}
// 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 N0;
}
// fold (sext_in_reg (setcc x)) -> setcc x iff (setcc x) == 0 or -1
- // FIXME: teach isSetCCEquivalent about 0, -1 and then use it here
if (N0.getOpcode() == ISD::SETCC &&
TLI.getSetCCResultContents() ==
TargetLowering::ZeroOrNegativeOneSetCCResult)
return N0;
- // FIXME: this code is currently just ported over from SelectionDAG.cpp
- // we probably actually want to handle this in two pieces. Rather than
- // checking all the top bits for zero, just check the sign bit here and turn
- // it into a zero extend inreg (AND with constant).
- // then, let the code for AND figure out if the mask is superfluous rather
- // than doing so here.
- if (N0.getOpcode() == ISD::AND &&
- N0.getOperand(1).getOpcode() == ISD::Constant) {
- uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
- unsigned NumBits = MVT::getSizeInBits(EVT);
- if ((Mask & (~0ULL << (NumBits-1))) == 0)
- return N0;
+ // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero
+ if (MaskedValueIsZero(N0, 1ULL << (EVTBits-1), TLI))
+ return DAG.getNode(ISD::AND, N0.getValueType(), N0,
+ DAG.getConstant(~0ULL >> (64-EVTBits), VT));
+ // fold (sext_in_reg (srl x)) -> sra x
+ if (N0.getOpcode() == ISD::SRL &&
+ N0.getOperand(1).getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(N0.getOperand(1))->getValue() == EVTBits) {
+ 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();
}
// and the truncate
return N0.getOperand(0);
}
+ // fold (truncate (load x)) -> (smaller load x)
+ 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();
+ // For big endian targets, we need to add an offset to the pointer to load
+ // the correct bytes. For little endian systems, we merely need to read
+ // fewer bytes from the same pointer.
+ uint64_t PtrOff =
+ (MVT::getSizeInBits(N0.getValueType()) - MVT::getSizeInBits(VT)) / 8;
+ SDOperand NewPtr = TLI.isLittleEndian() ? N0.getOperand(1) :
+ DAG.getNode(ISD::ADD, PtrType, N0.getOperand(1),
+ DAG.getConstant(PtrOff, PtrType));
+ WorkList.push_back(NewPtr.Val);
+ SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), NewPtr,N0.getOperand(2));
+ 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);
+
+ // 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);
+
+ // 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);
+
+ // 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();
+}
+
+SDOperand DAGCombiner::visitFDIV(SDNode *N) {
+ SDOperand N0 = N->getOperand(0);
+ SDOperand N1 = N->getOperand(1);
+ MVT::ValueType VT = N->getValueType(0);
+
+ 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(), VT);
+ }
+ return SDOperand();
+}
+
+SDOperand DAGCombiner::visitFREM(SDNode *N) {
+ SDOperand N0 = N->getOperand(0);
+ SDOperand N1 = N->getOperand(1);
+ MVT::ValueType VT = N->getValueType(0);
+
+ 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()), VT);
+ }
return SDOperand();
}
+
SDOperand DAGCombiner::visitSINT_TO_FP(SDNode *N) {
SDOperand N0 = N->getOperand(0);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
return SDOperand();
}
+SDOperand DAGCombiner::visitBRCOND(SDNode *N) {
+ SDOperand Chain = N->getOperand(0);
+ SDOperand N1 = N->getOperand(1);
+ SDOperand N2 = N->getOperand(2);
+ ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
+
+ // never taken branch, fold to chain
+ if (N1C && N1C->isNullValue())
+ return Chain;
+ // unconditional branch
+ if (N1C && N1C->getValue() == 1)
+ return DAG.getNode(ISD::BR, MVT::Other, Chain, N2);
+ return SDOperand();
+}
+
+SDOperand DAGCombiner::visitBRCONDTWOWAY(SDNode *N) {
+ SDOperand Chain = N->getOperand(0);
+ SDOperand N1 = N->getOperand(1);
+ SDOperand N2 = N->getOperand(2);
+ SDOperand N3 = N->getOperand(3);
+ ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
+
+ // unconditional branch to true mbb
+ if (N1C && N1C->getValue() == 1)
+ return DAG.getNode(ISD::BR, MVT::Other, Chain, N2);
+ // unconditional branch to false mbb
+ if (N1C && N1C->isNullValue())
+ return DAG.getNode(ISD::BR, MVT::Other, Chain, N3);
+ return SDOperand();
+}
+
+// Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
+//
+SDOperand DAGCombiner::visitBR_CC(SDNode *N) {
+ CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
+ SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
+
+ // Use SimplifySetCC to simplify SETCC's.
+ SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false);
+ ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(Simp.Val);
+
+ // fold br_cc true, dest -> br dest (unconditional branch)
+ if (SCCC && SCCC->getValue())
+ return DAG.getNode(ISD::BR, MVT::Other, N->getOperand(0),
+ N->getOperand(4));
+ // fold br_cc false, dest -> unconditional fall through
+ if (SCCC && SCCC->isNullValue())
+ return N->getOperand(0);
+ // fold to a simpler setcc
+ if (Simp.Val && Simp.getOpcode() == ISD::SETCC)
+ return DAG.getNode(ISD::BR_CC, MVT::Other, N->getOperand(0),
+ Simp.getOperand(2), Simp.getOperand(0),
+ Simp.getOperand(1), N->getOperand(4));
+ return SDOperand();
+}
+
+SDOperand DAGCombiner::visitBRTWOWAY_CC(SDNode *N) {
+ SDOperand Chain = N->getOperand(0);
+ SDOperand CCN = N->getOperand(1);
+ SDOperand LHS = N->getOperand(2);
+ SDOperand RHS = N->getOperand(3);
+ SDOperand N4 = N->getOperand(4);
+ SDOperand N5 = N->getOperand(5);
+
+ SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), LHS, RHS,
+ cast<CondCodeSDNode>(CCN)->get(), false);
+ ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
+
+ // fold select_cc lhs, rhs, x, x, cc -> x
+ if (N4 == N5)
+ return DAG.getNode(ISD::BR, MVT::Other, Chain, N4);
+ // fold select_cc true, x, y -> x
+ if (SCCC && SCCC->getValue())
+ return DAG.getNode(ISD::BR, MVT::Other, Chain, N4);
+ // fold select_cc false, x, y -> y
+ if (SCCC && SCCC->isNullValue())
+ return DAG.getNode(ISD::BR, MVT::Other, Chain, N5);
+ // fold to a simpler setcc
+ if (SCC.Val && SCC.getOpcode() == ISD::SETCC)
+ return DAG.getBR2Way_CC(Chain, SCC.getOperand(2), SCC.getOperand(0),
+ SCC.getOperand(1), N4, N5);
+ return SDOperand();
+}
+
+SDOperand DAGCombiner::visitLOAD(SDNode *N) {
+ SDOperand Chain = N->getOperand(0);
+ SDOperand Ptr = N->getOperand(1);
+ SDOperand SrcValue = N->getOperand(2);
+
+ // If this load is directly stored, replace the load value with the stored
+ // value.
+ // TODO: Handle store large -> read small portion.
+ // TODO: Handle TRUNCSTORE/EXTLOAD
+ if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr &&
+ Chain.getOperand(1).getValueType() == N->getValueType(0))
+ return CombineTo(N, Chain.getOperand(1), Chain);
+
+ return SDOperand();
+}
+
+SDOperand DAGCombiner::visitSTORE(SDNode *N) {
+ SDOperand Chain = N->getOperand(0);
+ SDOperand Value = N->getOperand(1);
+ SDOperand Ptr = N->getOperand(2);
+ SDOperand SrcValue = N->getOperand(3);
+
+ // If this is a store that kills a previous store, remove the previous store.
+ if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr &&
+ Chain.Val->hasOneUse() /* Avoid introducing DAG cycles */) {
+ // Create a new store of Value that replaces both stores.
+ SDNode *PrevStore = Chain.Val;
+ if (PrevStore->getOperand(1) == Value) // Same value multiply stored.
+ return Chain;
+ SDOperand NewStore = DAG.getNode(ISD::STORE, MVT::Other,
+ PrevStore->getOperand(0), Value, Ptr,
+ SrcValue);
+ CombineTo(N, NewStore); // Nuke this store.
+ CombineTo(PrevStore, NewStore); // Nuke the previous store.
+ return SDOperand(N, 0);
+ }
+
+ return SDOperand();
+}
+
+SDOperand DAGCombiner::SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2){
+ assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
+
+ SDOperand SCC = SimplifySelectCC(N0.getOperand(0), N0.getOperand(1), N1, N2,
+ cast<CondCodeSDNode>(N0.getOperand(2))->get());
+ // If we got a simplified select_cc node back from SimplifySelectCC, then
+ // break it down into a new SETCC node, and a new SELECT node, and then return
+ // the SELECT node, since we were called with a SELECT node.
+ if (SCC.Val) {
+ // Check to see if we got a select_cc back (to turn into setcc/select).
+ // Otherwise, just return whatever node we got back, like fabs.
+ if (SCC.getOpcode() == ISD::SELECT_CC) {
+ SDOperand SETCC = DAG.getNode(ISD::SETCC, N0.getValueType(),
+ SCC.getOperand(0), SCC.getOperand(1),
+ SCC.getOperand(4));
+ WorkList.push_back(SETCC.Val);
+ return DAG.getNode(ISD::SELECT, SCC.getValueType(), SCC.getOperand(2),
+ SCC.getOperand(3), SETCC);
+ }
+ return SCC;
+ }
+ 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) {
+
+ MVT::ValueType VT = N2.getValueType();
+ ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
+ ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
+ ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.Val);
+ ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.Val);
+
+ // Determine if the condition we're dealing with is constant
+ SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
+ ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
+
+ // fold select_cc true, x, y -> x
+ if (SCCC && SCCC->getValue())
+ return N2;
+ // fold select_cc false, x, y -> y
+ if (SCCC && SCCC->getValue() == 0)
+ return N3;
+
+ // Check to see if we can simplify the select into an fabs node
+ if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
+ // Allow either -0.0 or 0.0
+ if (CFP->getValue() == 0.0) {
+ // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
+ if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
+ N0 == N2 && N3.getOpcode() == ISD::FNEG &&
+ N2 == N3.getOperand(0))
+ return DAG.getNode(ISD::FABS, VT, N0);
+
+ // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
+ if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
+ N0 == N3 && N2.getOpcode() == ISD::FNEG &&
+ N2.getOperand(0) == N3)
+ return DAG.getNode(ISD::FABS, VT, N3);
+ }
+ }
+
+ // Check to see if we can perform the "gzip trick", transforming
+ // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A
+ if (N1C && N1C->isNullValue() && N3C && N3C->isNullValue() &&
+ MVT::isInteger(N0.getValueType()) &&
+ MVT::isInteger(N2.getValueType()) && CC == ISD::SETLT) {
+ MVT::ValueType XType = N0.getValueType();
+ MVT::ValueType AType = N2.getValueType();
+ if (XType >= AType) {
+ // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
+ // single-bit constant.
+ if (N2C && ((N2C->getValue() & (N2C->getValue()-1)) == 0)) {
+ unsigned ShCtV = Log2_64(N2C->getValue());
+ ShCtV = MVT::getSizeInBits(XType)-ShCtV-1;
+ SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy());
+ SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt);
+ WorkList.push_back(Shift.Val);
+ if (XType > AType) {
+ Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
+ WorkList.push_back(Shift.Val);
+ }
+ return DAG.getNode(ISD::AND, AType, Shift, N2);
+ }
+ SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
+ DAG.getConstant(MVT::getSizeInBits(XType)-1,
+ TLI.getShiftAmountTy()));
+ WorkList.push_back(Shift.Val);
+ if (XType > AType) {
+ Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
+ WorkList.push_back(Shift.Val);
+ }
+ return DAG.getNode(ISD::AND, AType, Shift, N2);
+ }
+ }
+
+ // fold select C, 16, 0 -> shl C, 4
+ if (N2C && N3C && N3C->isNullValue() && isPowerOf2_64(N2C->getValue()) &&
+ TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult) {
+ // Get a SetCC of the condition
+ // FIXME: Should probably make sure that setcc is legal if we ever have a
+ // target where it isn't.
+ SDOperand Temp, SCC = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
+ WorkList.push_back(SCC.Val);
+ // cast from setcc result type to select result type
+ if (AfterLegalize)
+ Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType());
+ else
+ Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC);
+ WorkList.push_back(Temp.Val);
+ // shl setcc result by log2 n2c
+ return DAG.getNode(ISD::SHL, N2.getValueType(), Temp,
+ DAG.getConstant(Log2_64(N2C->getValue()),
+ TLI.getShiftAmountTy()));
+ }
+
+ // Check to see if this is the equivalent of setcc
+ // FIXME: Turn all of these into setcc if setcc if setcc is legal
+ // otherwise, go ahead with the folds.
+ if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getValue() == 1ULL)) {
+ MVT::ValueType XType = N0.getValueType();
+ if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultTy())) {
+ SDOperand Res = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
+ if (Res.getValueType() != VT)
+ Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res);
+ return Res;
+ }
+
+ // seteq X, 0 -> srl (ctlz X, log2(size(X)))
+ if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
+ TLI.isOperationLegal(ISD::CTLZ, XType)) {
+ SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0);
+ return DAG.getNode(ISD::SRL, XType, Ctlz,
+ DAG.getConstant(Log2_32(MVT::getSizeInBits(XType)),
+ TLI.getShiftAmountTy()));
+ }
+ // setgt X, 0 -> srl (and (-X, ~X), size(X)-1)
+ if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
+ SDOperand NegN0 = DAG.getNode(ISD::SUB, XType, DAG.getConstant(0, XType),
+ N0);
+ SDOperand NotN0 = DAG.getNode(ISD::XOR, XType, N0,
+ DAG.getConstant(~0ULL, XType));
+ return DAG.getNode(ISD::SRL, XType,
+ DAG.getNode(ISD::AND, XType, NegN0, NotN0),
+ DAG.getConstant(MVT::getSizeInBits(XType)-1,
+ TLI.getShiftAmountTy()));
+ }
+ // setgt X, -1 -> xor (srl (X, size(X)-1), 1)
+ if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
+ SDOperand Sign = DAG.getNode(ISD::SRL, XType, N0,
+ DAG.getConstant(MVT::getSizeInBits(XType)-1,
+ TLI.getShiftAmountTy()));
+ return DAG.getNode(ISD::XOR, XType, Sign, DAG.getConstant(1, XType));
+ }
+ }
+
+ // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X ->
+ // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
+ if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) &&
+ N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) {
+ if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0))) {
+ MVT::ValueType XType = N0.getValueType();
+ if (SubC->isNullValue() && MVT::isInteger(XType)) {
+ SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
+ DAG.getConstant(MVT::getSizeInBits(XType)-1,
+ TLI.getShiftAmountTy()));
+ SDOperand Add = DAG.getNode(ISD::ADD, XType, N0, Shift);
+ WorkList.push_back(Shift.Val);
+ WorkList.push_back(Add.Val);
+ return DAG.getNode(ISD::XOR, XType, Add, Shift);
+ }
+ }
+ }
+
+ return SDOperand();
+}
+
SDOperand DAGCombiner::SimplifySetCC(MVT::ValueType VT, SDOperand N0,
- SDOperand N1, ISD::CondCode Cond) {
+ SDOperand N1, ISD::CondCode Cond,
+ bool foldBooleans) {
// These setcc operations always fold.
switch (Cond) {
default: break;
ExtDstTy),
Cond);
}
-
+
uint64_t MinVal, MaxVal;
unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
if (ISD::isSignedIntSetCC(Cond)) {
}
}
+ // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. Common for condcodes.
+ if (N0.getOpcode() == ISD::XOR)
+ if (ConstantSDNode *XORC = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
+ if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
+ // If we know that all of the inverted bits are zero, don't bother
+ // performing the inversion.
+ if (MaskedValueIsZero(N0.getOperand(0), ~XORC->getValue(), TLI))
+ return DAG.getSetCC(VT, N0.getOperand(0),
+ DAG.getConstant(XORC->getValue()^RHSC->getValue(),
+ N0.getValueType()), Cond);
+ }
+
// Simplify (X+Z) == X --> Z == 0
if (N0.getOperand(0) == N1)
return DAG.getSetCC(VT, N0.getOperand(1),
// Fold away ALL boolean setcc's.
SDOperand Temp;
- if (N0.getValueType() == MVT::i1) {
+ if (N0.getValueType() == MVT::i1 && foldBooleans) {
switch (Cond) {
default: assert(0 && "Unknown integer setcc!");
case ISD::SETEQ: // X == Y -> (X^Y)^1
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) {