/// make the right decision when generating code for different targets.
const X86Subtarget *Subtarget;
- /// CurBB - Current BB being isel'd.
- ///
- MachineBasicBlock *CurBB;
-
/// OptForSize - If true, selector should try to optimize for code size
/// instead of performance.
bool OptForSize;
private:
SDNode *Select(SDValue N);
SDNode *SelectAtomic64(SDNode *Node, unsigned Opc);
+ SDNode *SelectAtomicLoadAdd(SDNode *Node, MVT NVT);
bool MatchSegmentBaseAddress(SDValue N, X86ISelAddressMode &AM);
bool MatchLoad(SDValue N, X86ISelAddressMode &AM);
bool MatchWrapper(SDValue N, X86ISelAddressMode &AM);
- bool MatchAddress(SDValue N, X86ISelAddressMode &AM,
- unsigned Depth = 0);
+ bool MatchAddress(SDValue N, X86ISelAddressMode &AM);
+ bool MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
+ unsigned Depth);
bool MatchAddressBase(SDValue N, X86ISelAddressMode &AM);
bool SelectAddr(SDValue Op, SDValue N, SDValue &Base,
SDValue &Scale, SDValue &Index, SDValue &Disp,
/// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
/// when it has created a SelectionDAG for us to codegen.
void X86DAGToDAGISel::InstructionSelect() {
- CurBB = BB; // BB can change as result of isel.
- const Function *F = CurDAG->getMachineFunction().getFunction();
+ const Function *F = MF->getFunction();
OptForSize = F->hasFnAttr(Attribute::OptimizeForSize);
DEBUG(BB->dump());
// Codegen the basic block.
#ifndef NDEBUG
- DOUT << "===== Instruction selection begins:\n";
+ DEBUG(errs() << "===== Instruction selection begins:\n");
Indent = 0;
#endif
SelectRoot(*CurDAG);
#ifndef NDEBUG
- DOUT << "===== Instruction selection ends:\n";
+ DEBUG(errs() << "===== Instruction selection ends:\n");
#endif
CurDAG->RemoveDeadNodes();
/// MatchAddress - Add the specified node to the specified addressing mode,
/// returning true if it cannot be done. This just pattern matches for the
/// addressing mode.
-bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM,
- unsigned Depth) {
+bool X86DAGToDAGISel::MatchAddress(SDValue N, X86ISelAddressMode &AM) {
+ if (MatchAddressRecursively(N, AM, 0))
+ return true;
+
+ // Post-processing: Convert lea(,%reg,2) to lea(%reg,%reg), which has
+ // a smaller encoding and avoids a scaled-index.
+ if (AM.Scale == 2 &&
+ AM.BaseType == X86ISelAddressMode::RegBase &&
+ AM.Base.Reg.getNode() == 0) {
+ AM.Base.Reg = AM.IndexReg;
+ AM.Scale = 1;
+ }
+
+ return false;
+}
+
+bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
+ unsigned Depth) {
bool is64Bit = Subtarget->is64Bit();
DebugLoc dl = N.getDebugLoc();
- DOUT << "MatchAddress: "; DEBUG(AM.dump());
+ DEBUG(errs() << "MatchAddress: "); DEBUG(AM.dump());
// Limit recursion.
if (Depth > 5)
return MatchAddressBase(N, AM);
if (ConstantSDNode
*CN = dyn_cast<ConstantSDNode>(N.getNode()->getOperand(1))) {
unsigned Val = CN->getZExtValue();
+ // Note that we handle x<<1 as (,x,2) rather than (x,x) here so
+ // that the base operand remains free for further matching. If
+ // the base doesn't end up getting used, a post-processing step
+ // in MatchAddress turns (,x,2) into (x,x), which is cheaper.
if (Val == 1 || Val == 2 || Val == 3) {
AM.Scale = 1 << Val;
SDValue ShVal = N.getNode()->getOperand(0);
// Test if the LHS of the sub can be folded.
X86ISelAddressMode Backup = AM;
- if (MatchAddress(N.getNode()->getOperand(0), AM, Depth+1)) {
+ if (MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1)) {
AM = Backup;
break;
}
case ISD::ADD: {
X86ISelAddressMode Backup = AM;
- if (!MatchAddress(N.getNode()->getOperand(0), AM, Depth+1) &&
- !MatchAddress(N.getNode()->getOperand(1), AM, Depth+1))
+ if (!MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1) &&
+ !MatchAddressRecursively(N.getNode()->getOperand(1), AM, Depth+1))
return false;
AM = Backup;
- if (!MatchAddress(N.getNode()->getOperand(1), AM, Depth+1) &&
- !MatchAddress(N.getNode()->getOperand(0), AM, Depth+1))
+ if (!MatchAddressRecursively(N.getNode()->getOperand(1), AM, Depth+1) &&
+ !MatchAddressRecursively(N.getNode()->getOperand(0), AM, Depth+1))
return false;
AM = Backup;
X86ISelAddressMode Backup = AM;
uint64_t Offset = CN->getSExtValue();
// Start with the LHS as an addr mode.
- if (!MatchAddress(N.getOperand(0), AM, Depth+1) &&
+ if (!MatchAddressRecursively(N.getOperand(0), AM, Depth+1) &&
// Address could not have picked a GV address for the displacement.
AM.GV == NULL &&
// On x86-64, the resultant disp must fit in 32-bits.
/// initialize the global base register, if necessary.
///
SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
- MachineFunction *MF = CurBB->getParent();
unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
return CurDAG->getRegister(GlobalBaseReg, TLI.getPointerTy()).getNode();
}
array_lengthof(Ops));
}
+SDNode *X86DAGToDAGISel::SelectAtomicLoadAdd(SDNode *Node, MVT NVT) {
+ if (Node->hasAnyUseOfValue(0))
+ return 0;
+
+ // Optimize common patterns for __sync_add_and_fetch and
+ // __sync_sub_and_fetch where the result is not used. This allows us
+ // to use "lock" version of add, sub, inc, dec instructions.
+ // FIXME: Do not use special instructions but instead add the "lock"
+ // prefix to the target node somehow. The extra information will then be
+ // transferred to machine instruction and it denotes the prefix.
+ SDValue Chain = Node->getOperand(0);
+ SDValue Ptr = Node->getOperand(1);
+ SDValue Val = Node->getOperand(2);
+ SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
+ if (!SelectAddr(Ptr, Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
+ return 0;
+
+ bool isInc = false, isDec = false, isSub = false, isCN = false;
+ ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val);
+ if (CN) {
+ isCN = true;
+ int64_t CNVal = CN->getSExtValue();
+ if (CNVal == 1)
+ isInc = true;
+ else if (CNVal == -1)
+ isDec = true;
+ else if (CNVal >= 0)
+ Val = CurDAG->getTargetConstant(CNVal, NVT);
+ else {
+ isSub = true;
+ Val = CurDAG->getTargetConstant(-CNVal, NVT);
+ }
+ } else if (Val.hasOneUse() &&
+ Val.getOpcode() == ISD::SUB &&
+ X86::isZeroNode(Val.getOperand(0))) {
+ isSub = true;
+ Val = Val.getOperand(1);
+ }
+
+ unsigned Opc = 0;
+ switch (NVT.getSimpleVT()) {
+ default: return 0;
+ case MVT::i8:
+ if (isInc)
+ Opc = X86::LOCK_INC8m;
+ else if (isDec)
+ Opc = X86::LOCK_DEC8m;
+ else if (isSub) {
+ if (isCN)
+ Opc = X86::LOCK_SUB8mi;
+ else
+ Opc = X86::LOCK_SUB8mr;
+ } else {
+ if (isCN)
+ Opc = X86::LOCK_ADD8mi;
+ else
+ Opc = X86::LOCK_ADD8mr;
+ }
+ break;
+ case MVT::i16:
+ if (isInc)
+ Opc = X86::LOCK_INC16m;
+ else if (isDec)
+ Opc = X86::LOCK_DEC16m;
+ else if (isSub) {
+ if (isCN) {
+ if (Predicate_i16immSExt8(Val.getNode()))
+ Opc = X86::LOCK_SUB16mi8;
+ else
+ Opc = X86::LOCK_SUB16mi;
+ } else
+ Opc = X86::LOCK_SUB16mr;
+ } else {
+ if (isCN) {
+ if (Predicate_i16immSExt8(Val.getNode()))
+ Opc = X86::LOCK_ADD16mi8;
+ else
+ Opc = X86::LOCK_ADD16mi;
+ } else
+ Opc = X86::LOCK_ADD16mr;
+ }
+ break;
+ case MVT::i32:
+ if (isInc)
+ Opc = X86::LOCK_INC32m;
+ else if (isDec)
+ Opc = X86::LOCK_DEC32m;
+ else if (isSub) {
+ if (isCN) {
+ if (Predicate_i32immSExt8(Val.getNode()))
+ Opc = X86::LOCK_SUB32mi8;
+ else
+ Opc = X86::LOCK_SUB32mi;
+ } else
+ Opc = X86::LOCK_SUB32mr;
+ } else {
+ if (isCN) {
+ if (Predicate_i32immSExt8(Val.getNode()))
+ Opc = X86::LOCK_ADD32mi8;
+ else
+ Opc = X86::LOCK_ADD32mi;
+ } else
+ Opc = X86::LOCK_ADD32mr;
+ }
+ break;
+ case MVT::i64:
+ if (isInc)
+ Opc = X86::LOCK_INC64m;
+ else if (isDec)
+ Opc = X86::LOCK_DEC64m;
+ else if (isSub) {
+ Opc = X86::LOCK_SUB64mr;
+ if (isCN) {
+ if (Predicate_i64immSExt8(Val.getNode()))
+ Opc = X86::LOCK_SUB64mi8;
+ else if (Predicate_i64immSExt32(Val.getNode()))
+ Opc = X86::LOCK_SUB64mi32;
+ }
+ } else {
+ Opc = X86::LOCK_ADD64mr;
+ if (isCN) {
+ if (Predicate_i64immSExt8(Val.getNode()))
+ Opc = X86::LOCK_ADD64mi8;
+ else if (Predicate_i64immSExt32(Val.getNode()))
+ Opc = X86::LOCK_ADD64mi32;
+ }
+ }
+ break;
+ }
+
+ DebugLoc dl = Node->getDebugLoc();
+ SDValue Undef = SDValue(CurDAG->getTargetNode(TargetInstrInfo::IMPLICIT_DEF,
+ dl, NVT), 0);
+ SDValue MemOp = CurDAG->getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
+ if (isInc || isDec) {
+ SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, MemOp, Chain };
+ SDValue Ret = SDValue(CurDAG->getTargetNode(Opc, dl, MVT::Other, Ops, 7), 0);
+ SDValue RetVals[] = { Undef, Ret };
+ return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
+ } else {
+ SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, MemOp, Chain };
+ SDValue Ret = SDValue(CurDAG->getTargetNode(Opc, dl, MVT::Other, Ops, 8), 0);
+ SDValue RetVals[] = { Undef, Ret };
+ return CurDAG->getMergeValues(RetVals, 2, dl).getNode();
+ }
+}
+
SDNode *X86DAGToDAGISel::Select(SDValue N) {
SDNode *Node = N.getNode();
MVT NVT = Node->getValueType(0);
DebugLoc dl = Node->getDebugLoc();
#ifndef NDEBUG
- DOUT << std::string(Indent, ' ') << "Selecting: ";
+ DEBUG(errs() << std::string(Indent, ' ') << "Selecting: ");
DEBUG(Node->dump(CurDAG));
- DOUT << "\n";
+ DEBUG(errs() << "\n");
Indent += 2;
#endif
if (Node->isMachineOpcode()) {
#ifndef NDEBUG
- DOUT << std::string(Indent-2, ' ') << "== ";
+ DEBUG(errs() << std::string(Indent-2, ' ') << "== ");
DEBUG(Node->dump(CurDAG));
- DOUT << "\n";
+ DEBUG(errs() << "\n");
Indent -= 2;
#endif
return NULL; // Already selected.
}
switch (Opcode) {
- default: break;
- case X86ISD::GlobalBaseReg:
- return getGlobalBaseReg();
-
- case X86ISD::ATOMOR64_DAG:
- return SelectAtomic64(Node, X86::ATOMOR6432);
- case X86ISD::ATOMXOR64_DAG:
- return SelectAtomic64(Node, X86::ATOMXOR6432);
- case X86ISD::ATOMADD64_DAG:
- return SelectAtomic64(Node, X86::ATOMADD6432);
- case X86ISD::ATOMSUB64_DAG:
- return SelectAtomic64(Node, X86::ATOMSUB6432);
- case X86ISD::ATOMNAND64_DAG:
- return SelectAtomic64(Node, X86::ATOMNAND6432);
- case X86ISD::ATOMAND64_DAG:
- return SelectAtomic64(Node, X86::ATOMAND6432);
- case X86ISD::ATOMSWAP64_DAG:
- return SelectAtomic64(Node, X86::ATOMSWAP6432);
-
- case ISD::SMUL_LOHI:
- case ISD::UMUL_LOHI: {
- SDValue N0 = Node->getOperand(0);
- SDValue N1 = Node->getOperand(1);
-
- bool isSigned = Opcode == ISD::SMUL_LOHI;
- if (!isSigned)
- switch (NVT.getSimpleVT()) {
- default: LLVM_UNREACHABLE("Unsupported VT!");
- case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
- case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
- case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
- case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
- }
- else
- switch (NVT.getSimpleVT()) {
- default: LLVM_UNREACHABLE("Unsupported VT!");
- case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
- case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
- case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
- case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
- }
+ default: break;
+ case X86ISD::GlobalBaseReg:
+ return getGlobalBaseReg();
+
+ case X86ISD::ATOMOR64_DAG:
+ return SelectAtomic64(Node, X86::ATOMOR6432);
+ case X86ISD::ATOMXOR64_DAG:
+ return SelectAtomic64(Node, X86::ATOMXOR6432);
+ case X86ISD::ATOMADD64_DAG:
+ return SelectAtomic64(Node, X86::ATOMADD6432);
+ case X86ISD::ATOMSUB64_DAG:
+ return SelectAtomic64(Node, X86::ATOMSUB6432);
+ case X86ISD::ATOMNAND64_DAG:
+ return SelectAtomic64(Node, X86::ATOMNAND6432);
+ case X86ISD::ATOMAND64_DAG:
+ return SelectAtomic64(Node, X86::ATOMAND6432);
+ case X86ISD::ATOMSWAP64_DAG:
+ return SelectAtomic64(Node, X86::ATOMSWAP6432);
+
+ case ISD::ATOMIC_LOAD_ADD: {
+ SDNode *RetVal = SelectAtomicLoadAdd(Node, NVT);
+ if (RetVal)
+ return RetVal;
+ break;
+ }
- unsigned LoReg, HiReg;
+ case ISD::SMUL_LOHI:
+ case ISD::UMUL_LOHI: {
+ SDValue N0 = Node->getOperand(0);
+ SDValue N1 = Node->getOperand(1);
+
+ bool isSigned = Opcode == ISD::SMUL_LOHI;
+ if (!isSigned)
switch (NVT.getSimpleVT()) {
- default: LLVM_UNREACHABLE("Unsupported VT!");
- case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
- case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
- case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
- case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
+ default: llvm_unreachable("Unsupported VT!");
+ case MVT::i8: Opc = X86::MUL8r; MOpc = X86::MUL8m; break;
+ case MVT::i16: Opc = X86::MUL16r; MOpc = X86::MUL16m; break;
+ case MVT::i32: Opc = X86::MUL32r; MOpc = X86::MUL32m; break;
+ case MVT::i64: Opc = X86::MUL64r; MOpc = X86::MUL64m; break;
}
-
- SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
- bool foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
- // multiplty is commmutative
- if (!foldedLoad) {
- foldedLoad = TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
- if (foldedLoad)
- std::swap(N0, N1);
+ else
+ switch (NVT.getSimpleVT()) {
+ default: llvm_unreachable("Unsupported VT!");
+ case MVT::i8: Opc = X86::IMUL8r; MOpc = X86::IMUL8m; break;
+ case MVT::i16: Opc = X86::IMUL16r; MOpc = X86::IMUL16m; break;
+ case MVT::i32: Opc = X86::IMUL32r; MOpc = X86::IMUL32m; break;
+ case MVT::i64: Opc = X86::IMUL64r; MOpc = X86::IMUL64m; break;
}
- SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
- N0, SDValue()).getValue(1);
-
- if (foldedLoad) {
- SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
- InFlag };
- SDNode *CNode =
- CurDAG->getTargetNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
- array_lengthof(Ops));
- InFlag = SDValue(CNode, 1);
- // Update the chain.
- ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
- } else {
- InFlag =
- SDValue(CurDAG->getTargetNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
- }
+ unsigned LoReg, HiReg;
+ switch (NVT.getSimpleVT()) {
+ default: llvm_unreachable("Unsupported VT!");
+ case MVT::i8: LoReg = X86::AL; HiReg = X86::AH; break;
+ case MVT::i16: LoReg = X86::AX; HiReg = X86::DX; break;
+ case MVT::i32: LoReg = X86::EAX; HiReg = X86::EDX; break;
+ case MVT::i64: LoReg = X86::RAX; HiReg = X86::RDX; break;
+ }
- // Copy the low half of the result, if it is needed.
- if (!N.getValue(0).use_empty()) {
- SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
- LoReg, NVT, InFlag);
- InFlag = Result.getValue(2);
- ReplaceUses(N.getValue(0), Result);
+ SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
+ bool foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
+ // multiplty is commmutative
+ if (!foldedLoad) {
+ foldedLoad = TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
+ if (foldedLoad)
+ std::swap(N0, N1);
+ }
+
+ SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, LoReg,
+ N0, SDValue()).getValue(1);
+
+ if (foldedLoad) {
+ SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
+ InFlag };
+ SDNode *CNode =
+ CurDAG->getTargetNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
+ array_lengthof(Ops));
+ InFlag = SDValue(CNode, 1);
+ // Update the chain.
+ ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
+ } else {
+ InFlag =
+ SDValue(CurDAG->getTargetNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
+ }
+
+ // Copy the low half of the result, if it is needed.
+ if (!N.getValue(0).use_empty()) {
+ SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+ LoReg, NVT, InFlag);
+ InFlag = Result.getValue(2);
+ ReplaceUses(N.getValue(0), Result);
#ifndef NDEBUG
- DOUT << std::string(Indent-2, ' ') << "=> ";
- DEBUG(Result.getNode()->dump(CurDAG));
- DOUT << "\n";
+ DEBUG(errs() << std::string(Indent-2, ' ') << "=> ");
+ DEBUG(Result.getNode()->dump(CurDAG));
+ DEBUG(errs() << "\n");
#endif
+ }
+ // Copy the high half of the result, if it is needed.
+ if (!N.getValue(1).use_empty()) {
+ SDValue Result;
+ if (HiReg == X86::AH && Subtarget->is64Bit()) {
+ // Prevent use of AH in a REX instruction by referencing AX instead.
+ // Shift it down 8 bits.
+ Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+ X86::AX, MVT::i16, InFlag);
+ InFlag = Result.getValue(2);
+ Result = SDValue(CurDAG->getTargetNode(X86::SHR16ri, dl, MVT::i16,
+ Result,
+ CurDAG->getTargetConstant(8, MVT::i8)), 0);
+ // Then truncate it down to i8.
+ SDValue SRIdx = CurDAG->getTargetConstant(X86::SUBREG_8BIT, MVT::i32);
+ Result = SDValue(CurDAG->getTargetNode(X86::EXTRACT_SUBREG, dl,
+ MVT::i8, Result, SRIdx), 0);
+ } else {
+ Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+ HiReg, NVT, InFlag);
+ InFlag = Result.getValue(2);
}
- // Copy the high half of the result, if it is needed.
- if (!N.getValue(1).use_empty()) {
- SDValue Result;
- if (HiReg == X86::AH && Subtarget->is64Bit()) {
- // Prevent use of AH in a REX instruction by referencing AX instead.
- // Shift it down 8 bits.
- Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
- X86::AX, MVT::i16, InFlag);
- InFlag = Result.getValue(2);
- Result = SDValue(CurDAG->getTargetNode(X86::SHR16ri, dl, MVT::i16,
- Result,
- CurDAG->getTargetConstant(8, MVT::i8)), 0);
- // Then truncate it down to i8.
- SDValue SRIdx = CurDAG->getTargetConstant(X86::SUBREG_8BIT, MVT::i32);
- Result = SDValue(CurDAG->getTargetNode(X86::EXTRACT_SUBREG, dl,
- MVT::i8, Result, SRIdx), 0);
- } else {
- Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
- HiReg, NVT, InFlag);
- InFlag = Result.getValue(2);
- }
- ReplaceUses(N.getValue(1), Result);
+ ReplaceUses(N.getValue(1), Result);
#ifndef NDEBUG
- DOUT << std::string(Indent-2, ' ') << "=> ";
- DEBUG(Result.getNode()->dump(CurDAG));
- DOUT << "\n";
+ DEBUG(errs() << std::string(Indent-2, ' ') << "=> ");
+ DEBUG(Result.getNode()->dump(CurDAG));
+ DEBUG(errs() << "\n");
#endif
- }
+ }
#ifndef NDEBUG
- Indent -= 2;
+ Indent -= 2;
#endif
- return NULL;
- }
-
- case ISD::SDIVREM:
- case ISD::UDIVREM: {
- SDValue N0 = Node->getOperand(0);
- SDValue N1 = Node->getOperand(1);
-
- bool isSigned = Opcode == ISD::SDIVREM;
- if (!isSigned)
- switch (NVT.getSimpleVT()) {
- default: LLVM_UNREACHABLE("Unsupported VT!");
- case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
- case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
- case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
- case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
- }
- else
- switch (NVT.getSimpleVT()) {
- default: LLVM_UNREACHABLE("Unsupported VT!");
- case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
- case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
- case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
- case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
- }
+ return NULL;
+ }
- unsigned LoReg, HiReg;
- unsigned ClrOpcode, SExtOpcode;
+ case ISD::SDIVREM:
+ case ISD::UDIVREM: {
+ SDValue N0 = Node->getOperand(0);
+ SDValue N1 = Node->getOperand(1);
+
+ bool isSigned = Opcode == ISD::SDIVREM;
+ if (!isSigned)
+ switch (NVT.getSimpleVT()) {
+ default: llvm_unreachable("Unsupported VT!");
+ case MVT::i8: Opc = X86::DIV8r; MOpc = X86::DIV8m; break;
+ case MVT::i16: Opc = X86::DIV16r; MOpc = X86::DIV16m; break;
+ case MVT::i32: Opc = X86::DIV32r; MOpc = X86::DIV32m; break;
+ case MVT::i64: Opc = X86::DIV64r; MOpc = X86::DIV64m; break;
+ }
+ else
switch (NVT.getSimpleVT()) {
- default: LLVM_UNREACHABLE("Unsupported VT!");
- case MVT::i8:
- LoReg = X86::AL; HiReg = X86::AH;
- ClrOpcode = 0;
- SExtOpcode = X86::CBW;
- break;
- case MVT::i16:
- LoReg = X86::AX; HiReg = X86::DX;
- ClrOpcode = X86::MOV16r0;
- SExtOpcode = X86::CWD;
- break;
- case MVT::i32:
- LoReg = X86::EAX; HiReg = X86::EDX;
- ClrOpcode = X86::MOV32r0;
- SExtOpcode = X86::CDQ;
- break;
- case MVT::i64:
- LoReg = X86::RAX; HiReg = X86::RDX;
- ClrOpcode = X86::MOV64r0;
- SExtOpcode = X86::CQO;
- break;
+ default: llvm_unreachable("Unsupported VT!");
+ case MVT::i8: Opc = X86::IDIV8r; MOpc = X86::IDIV8m; break;
+ case MVT::i16: Opc = X86::IDIV16r; MOpc = X86::IDIV16m; break;
+ case MVT::i32: Opc = X86::IDIV32r; MOpc = X86::IDIV32m; break;
+ case MVT::i64: Opc = X86::IDIV64r; MOpc = X86::IDIV64m; break;
}
- SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
- bool foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
- bool signBitIsZero = CurDAG->SignBitIsZero(N0);
-
- SDValue InFlag;
- if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
- // Special case for div8, just use a move with zero extension to AX to
- // clear the upper 8 bits (AH).
- SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
- if (TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
- SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
- Move =
- SDValue(CurDAG->getTargetNode(X86::MOVZX16rm8, dl, MVT::i16,
- MVT::Other, Ops,
- array_lengthof(Ops)), 0);
- Chain = Move.getValue(1);
- ReplaceUses(N0.getValue(1), Chain);
- } else {
- Move =
- SDValue(CurDAG->getTargetNode(X86::MOVZX16rr8, dl, MVT::i16, N0),0);
- Chain = CurDAG->getEntryNode();
- }
- Chain = CurDAG->getCopyToReg(Chain, dl, X86::AX, Move, SDValue());
- InFlag = Chain.getValue(1);
+ unsigned LoReg, HiReg;
+ unsigned ClrOpcode, SExtOpcode;
+ switch (NVT.getSimpleVT()) {
+ default: llvm_unreachable("Unsupported VT!");
+ case MVT::i8:
+ LoReg = X86::AL; HiReg = X86::AH;
+ ClrOpcode = 0;
+ SExtOpcode = X86::CBW;
+ break;
+ case MVT::i16:
+ LoReg = X86::AX; HiReg = X86::DX;
+ ClrOpcode = X86::MOV16r0;
+ SExtOpcode = X86::CWD;
+ break;
+ case MVT::i32:
+ LoReg = X86::EAX; HiReg = X86::EDX;
+ ClrOpcode = X86::MOV32r0;
+ SExtOpcode = X86::CDQ;
+ break;
+ case MVT::i64:
+ LoReg = X86::RAX; HiReg = X86::RDX;
+ ClrOpcode = ~0U; // NOT USED.
+ SExtOpcode = X86::CQO;
+ break;
+ }
+
+ SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
+ bool foldedLoad = TryFoldLoad(N, N1, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4);
+ bool signBitIsZero = CurDAG->SignBitIsZero(N0);
+
+ SDValue InFlag;
+ if (NVT == MVT::i8 && (!isSigned || signBitIsZero)) {
+ // Special case for div8, just use a move with zero extension to AX to
+ // clear the upper 8 bits (AH).
+ SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Move, Chain;
+ if (TryFoldLoad(N, N0, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4)) {
+ SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N0.getOperand(0) };
+ Move =
+ SDValue(CurDAG->getTargetNode(X86::MOVZX16rm8, dl, MVT::i16,
+ MVT::Other, Ops,
+ array_lengthof(Ops)), 0);
+ Chain = Move.getValue(1);
+ ReplaceUses(N0.getValue(1), Chain);
} else {
+ Move =
+ SDValue(CurDAG->getTargetNode(X86::MOVZX16rr8, dl, MVT::i16, N0),0);
+ Chain = CurDAG->getEntryNode();
+ }
+ Chain = CurDAG->getCopyToReg(Chain, dl, X86::AX, Move, SDValue());
+ InFlag = Chain.getValue(1);
+ } else {
+ InFlag =
+ CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
+ LoReg, N0, SDValue()).getValue(1);
+ if (isSigned && !signBitIsZero) {
+ // Sign extend the low part into the high part.
InFlag =
- CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl,
- LoReg, N0, SDValue()).getValue(1);
- if (isSigned && !signBitIsZero) {
- // Sign extend the low part into the high part.
- InFlag =
- SDValue(CurDAG->getTargetNode(SExtOpcode, dl, MVT::Flag, InFlag),0);
+ SDValue(CurDAG->getTargetNode(SExtOpcode, dl, MVT::Flag, InFlag),0);
+ } else {
+ // Zero out the high part, effectively zero extending the input.
+ SDValue ClrNode;
+
+ if (NVT.getSimpleVT() == MVT::i64) {
+ ClrNode = SDValue(CurDAG->getTargetNode(X86::MOV32r0, dl, MVT::i32),
+ 0);
+ // We just did a 32-bit clear, insert it into a 64-bit register to
+ // clear the whole 64-bit reg.
+ SDValue Undef =
+ SDValue(CurDAG->getTargetNode(TargetInstrInfo::IMPLICIT_DEF,
+ dl, MVT::i64), 0);
+ SDValue SubRegNo =
+ CurDAG->getTargetConstant(X86::SUBREG_32BIT, MVT::i32);
+ ClrNode =
+ SDValue(CurDAG->getTargetNode(TargetInstrInfo::INSERT_SUBREG, dl,
+ MVT::i64, Undef, ClrNode, SubRegNo),
+ 0);
} else {
- // Zero out the high part, effectively zero extending the input.
- SDValue ClrNode = SDValue(CurDAG->getTargetNode(ClrOpcode, dl, NVT),
- 0);
- InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, HiReg,
- ClrNode, InFlag).getValue(1);
+ ClrNode = SDValue(CurDAG->getTargetNode(ClrOpcode, dl, NVT), 0);
}
- }
- if (foldedLoad) {
- SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
- InFlag };
- SDNode *CNode =
- CurDAG->getTargetNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
- array_lengthof(Ops));
- InFlag = SDValue(CNode, 1);
- // Update the chain.
- ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
- } else {
- InFlag =
- SDValue(CurDAG->getTargetNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
+ InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, HiReg,
+ ClrNode, InFlag).getValue(1);
}
+ }
- // Copy the division (low) result, if it is needed.
- if (!N.getValue(0).use_empty()) {
- SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
- LoReg, NVT, InFlag);
- InFlag = Result.getValue(2);
- ReplaceUses(N.getValue(0), Result);
+ if (foldedLoad) {
+ SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
+ InFlag };
+ SDNode *CNode =
+ CurDAG->getTargetNode(MOpc, dl, MVT::Other, MVT::Flag, Ops,
+ array_lengthof(Ops));
+ InFlag = SDValue(CNode, 1);
+ // Update the chain.
+ ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
+ } else {
+ InFlag =
+ SDValue(CurDAG->getTargetNode(Opc, dl, MVT::Flag, N1, InFlag), 0);
+ }
+
+ // Copy the division (low) result, if it is needed.
+ if (!N.getValue(0).use_empty()) {
+ SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+ LoReg, NVT, InFlag);
+ InFlag = Result.getValue(2);
+ ReplaceUses(N.getValue(0), Result);
#ifndef NDEBUG
- DOUT << std::string(Indent-2, ' ') << "=> ";
- DEBUG(Result.getNode()->dump(CurDAG));
- DOUT << "\n";
+ DEBUG(errs() << std::string(Indent-2, ' ') << "=> ");
+ DEBUG(Result.getNode()->dump(CurDAG));
+ DEBUG(errs() << "\n");
#endif
+ }
+ // Copy the remainder (high) result, if it is needed.
+ if (!N.getValue(1).use_empty()) {
+ SDValue Result;
+ if (HiReg == X86::AH && Subtarget->is64Bit()) {
+ // Prevent use of AH in a REX instruction by referencing AX instead.
+ // Shift it down 8 bits.
+ Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+ X86::AX, MVT::i16, InFlag);
+ InFlag = Result.getValue(2);
+ Result = SDValue(CurDAG->getTargetNode(X86::SHR16ri, dl, MVT::i16,
+ Result,
+ CurDAG->getTargetConstant(8, MVT::i8)),
+ 0);
+ // Then truncate it down to i8.
+ SDValue SRIdx = CurDAG->getTargetConstant(X86::SUBREG_8BIT, MVT::i32);
+ Result = SDValue(CurDAG->getTargetNode(X86::EXTRACT_SUBREG, dl,
+ MVT::i8, Result, SRIdx), 0);
+ } else {
+ Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
+ HiReg, NVT, InFlag);
+ InFlag = Result.getValue(2);
}
- // Copy the remainder (high) result, if it is needed.
- if (!N.getValue(1).use_empty()) {
- SDValue Result;
- if (HiReg == X86::AH && Subtarget->is64Bit()) {
- // Prevent use of AH in a REX instruction by referencing AX instead.
- // Shift it down 8 bits.
- Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
- X86::AX, MVT::i16, InFlag);
- InFlag = Result.getValue(2);
- Result = SDValue(CurDAG->getTargetNode(X86::SHR16ri, dl, MVT::i16,
- Result,
- CurDAG->getTargetConstant(8, MVT::i8)),
- 0);
- // Then truncate it down to i8.
- SDValue SRIdx = CurDAG->getTargetConstant(X86::SUBREG_8BIT, MVT::i32);
- Result = SDValue(CurDAG->getTargetNode(X86::EXTRACT_SUBREG, dl,
- MVT::i8, Result, SRIdx), 0);
- } else {
- Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
- HiReg, NVT, InFlag);
- InFlag = Result.getValue(2);
- }
- ReplaceUses(N.getValue(1), Result);
+ ReplaceUses(N.getValue(1), Result);
#ifndef NDEBUG
- DOUT << std::string(Indent-2, ' ') << "=> ";
- DEBUG(Result.getNode()->dump(CurDAG));
- DOUT << "\n";
+ DEBUG(errs() << std::string(Indent-2, ' ') << "=> ");
+ DEBUG(Result.getNode()->dump(CurDAG));
+ DEBUG(errs() << "\n");
#endif
- }
+ }
#ifndef NDEBUG
- Indent -= 2;
+ Indent -= 2;
#endif
+ return NULL;
+ }
+
+ case ISD::DECLARE: {
+ // Handle DECLARE nodes here because the second operand may have been
+ // wrapped in X86ISD::Wrapper.
+ SDValue Chain = Node->getOperand(0);
+ SDValue N1 = Node->getOperand(1);
+ SDValue N2 = Node->getOperand(2);
+ FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N1);
+
+ // FIXME: We need to handle this for VLAs.
+ if (!FINode) {
+ ReplaceUses(N.getValue(0), Chain);
return NULL;
}
- case ISD::DECLARE: {
- // Handle DECLARE nodes here because the second operand may have been
- // wrapped in X86ISD::Wrapper.
- SDValue Chain = Node->getOperand(0);
- SDValue N1 = Node->getOperand(1);
- SDValue N2 = Node->getOperand(2);
- FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N1);
-
- // FIXME: We need to handle this for VLAs.
- if (!FINode) {
- ReplaceUses(N.getValue(0), Chain);
- return NULL;
- }
-
- if (N2.getOpcode() == ISD::ADD &&
- N2.getOperand(0).getOpcode() == X86ISD::GlobalBaseReg)
- N2 = N2.getOperand(1);
-
- // If N2 is not Wrapper(decriptor) then the llvm.declare is mangled
- // somehow, just ignore it.
- if (N2.getOpcode() != X86ISD::Wrapper &&
- N2.getOpcode() != X86ISD::WrapperRIP) {
- ReplaceUses(N.getValue(0), Chain);
- return NULL;
- }
- GlobalAddressSDNode *GVNode =
- dyn_cast<GlobalAddressSDNode>(N2.getOperand(0));
- if (GVNode == 0) {
- ReplaceUses(N.getValue(0), Chain);
- return NULL;
- }
- SDValue Tmp1 = CurDAG->getTargetFrameIndex(FINode->getIndex(),
- TLI.getPointerTy());
- SDValue Tmp2 = CurDAG->getTargetGlobalAddress(GVNode->getGlobal(),
- TLI.getPointerTy());
- SDValue Ops[] = { Tmp1, Tmp2, Chain };
- return CurDAG->getTargetNode(TargetInstrInfo::DECLARE, dl,
- MVT::Other, Ops,
- array_lengthof(Ops));
+ if (N2.getOpcode() == ISD::ADD &&
+ N2.getOperand(0).getOpcode() == X86ISD::GlobalBaseReg)
+ N2 = N2.getOperand(1);
+
+ // If N2 is not Wrapper(decriptor) then the llvm.declare is mangled
+ // somehow, just ignore it.
+ if (N2.getOpcode() != X86ISD::Wrapper &&
+ N2.getOpcode() != X86ISD::WrapperRIP) {
+ ReplaceUses(N.getValue(0), Chain);
+ return NULL;
}
+ GlobalAddressSDNode *GVNode =
+ dyn_cast<GlobalAddressSDNode>(N2.getOperand(0));
+ if (GVNode == 0) {
+ ReplaceUses(N.getValue(0), Chain);
+ return NULL;
+ }
+ SDValue Tmp1 = CurDAG->getTargetFrameIndex(FINode->getIndex(),
+ TLI.getPointerTy());
+ SDValue Tmp2 = CurDAG->getTargetGlobalAddress(GVNode->getGlobal(),
+ TLI.getPointerTy());
+ SDValue Ops[] = { Tmp1, Tmp2, Chain };
+ return CurDAG->getTargetNode(TargetInstrInfo::DECLARE, dl,
+ MVT::Other, Ops,
+ array_lengthof(Ops));
+ }
}
SDNode *ResNode = SelectCode(N);
#ifndef NDEBUG
- DOUT << std::string(Indent-2, ' ') << "=> ";
+ DEBUG(errs() << std::string(Indent-2, ' ') << "=> ");
if (ResNode == NULL || ResNode == N.getNode())
DEBUG(N.getNode()->dump(CurDAG));
else
DEBUG(ResNode->dump(CurDAG));
- DOUT << "\n";
+ DEBUG(errs() << "\n");
Indent -= 2;
#endif