///
//===---------------------------------------------------------------------===//
namespace {
- cl::opt<RegisterScheduler::FunctionPassCtor, false,
- RegisterPassParser<RegisterScheduler> >
+ static cl::opt<RegisterScheduler::FunctionPassCtor, false,
+ RegisterPassParser<RegisterScheduler> >
ISHeuristic("pre-RA-sched",
cl::init(&createDefaultScheduler),
cl::desc("Instruction schedulers available (before register"
RegsForValue() : TLI(0) {}
- RegsForValue(const TargetLowering &tli,
- unsigned Reg, MVT::ValueType regvt, MVT::ValueType valuevt)
- : TLI(&tli), ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(1, Reg) {}
RegsForValue(const TargetLowering &tli,
const SmallVector<unsigned, 4> ®s,
MVT::ValueType regvt, MVT::ValueType valuevt)
}
}
+ /// append - Add the specified values to this one.
+ void append(const RegsForValue &RHS) {
+ TLI = RHS.TLI;
+ ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
+ RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
+ Regs.append(RHS.Regs.begin(), RHS.Regs.end());
+ }
+
+
/// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
/// this value and returns the result as a ValueVTs value. This uses
/// Chain/Flag as the input and updates them for the output Chain/Flag.
assert(0 && "UserOp2 should not exist at instruction selection time!");
abort();
}
+
+private:
+ inline const char *implVisitBinaryAtomic(CallInst& I, ISD::NodeType Op);
+
};
} // end namespace llvm
case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break;
case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break;
case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break;
- case FCmpInst::FCMP_ORD: FOC = ISD::SETEQ; FPC = ISD::SETO; break;
- case FCmpInst::FCMP_UNO: FOC = ISD::SETNE; FPC = ISD::SETUO; break;
+ case FCmpInst::FCMP_ORD: FOC = FPC = ISD::SETO; break;
+ case FCmpInst::FCMP_UNO: FOC = FPC = ISD::SETUO; break;
case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break;
case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break;
case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break;
case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break;
case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break;
case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break;
- case FCmpInst::FCMP_ORD: FOC = ISD::SETEQ; FPC = ISD::SETO; break;
- case FCmpInst::FCMP_UNO: FOC = ISD::SETNE; FPC = ISD::SETUO; break;
+ case FCmpInst::FCMP_ORD: FOC = FPC = ISD::SETO; break;
+ case FCmpInst::FCMP_UNO: FOC = FPC = ISD::SETUO; break;
case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break;
case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break;
case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break;
/// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
static GlobalVariable *ExtractTypeInfo (Value *V) {
- V = IntrinsicInst::StripPointerCasts(V);
+ V = V->stripPointerCasts();
GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
assert ((GV || isa<ConstantPointerNull>(V)) &&
"TypeInfo must be a global variable or NULL");
}
}
+
+/// Inlined utility function to implement binary input atomic intrinsics for
+// visitIntrinsicCall: I is a call instruction
+// Op is the associated NodeType for I
+const char *
+SelectionDAGLowering::implVisitBinaryAtomic(CallInst& I, ISD::NodeType Op) {
+ SDOperand Root = getRoot();
+ SDOperand O2 = getValue(I.getOperand(2));
+ SDOperand L = DAG.getAtomic(Op, Root,
+ getValue(I.getOperand(1)),
+ O2, O2.getValueType());
+ setValue(&I, L);
+ DAG.setRoot(L.getValue(1));
+ return 0;
+}
+
/// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
/// we want to emit this as a call to a named external function, return the name
/// otherwise lower it and return null.
return 0;
case Intrinsic::init_trampoline: {
- const Function *F =
- cast<Function>(IntrinsicInst::StripPointerCasts(I.getOperand(2)));
+ const Function *F = cast<Function>(I.getOperand(2)->stripPointerCasts());
SDOperand Ops[6];
Ops[0] = getRoot();
DAG.setRoot(L.getValue(1));
return 0;
}
- case Intrinsic::atomic_las: {
- SDOperand Root = getRoot();
- SDOperand O2 = getValue(I.getOperand(2));
- SDOperand L = DAG.getAtomic(ISD::ATOMIC_LAS, Root,
- getValue(I.getOperand(1)),
- O2, O2.getValueType());
- setValue(&I, L);
- DAG.setRoot(L.getValue(1));
- return 0;
- }
- case Intrinsic::atomic_swap: {
- SDOperand Root = getRoot();
- SDOperand O2 = getValue(I.getOperand(2));
- SDOperand L = DAG.getAtomic(ISD::ATOMIC_SWAP, Root,
- getValue(I.getOperand(1)),
- O2, O2.getValueType());
- setValue(&I, L);
- DAG.setRoot(L.getValue(1));
- return 0;
- }
-
+ case Intrinsic::atomic_las:
+ return implVisitBinaryAtomic(I, ISD::ATOMIC_LAS);
+ case Intrinsic::atomic_lss:
+ return implVisitBinaryAtomic(I, ISD::ATOMIC_LSS);
+ case Intrinsic::atomic_load_and:
+ return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND);
+ case Intrinsic::atomic_load_or:
+ return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR);
+ case Intrinsic::atomic_load_xor:
+ return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR);
+ case Intrinsic::atomic_load_min:
+ return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN);
+ case Intrinsic::atomic_load_max:
+ return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX);
+ case Intrinsic::atomic_load_umin:
+ return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN);
+ case Intrinsic::atomic_load_umax:
+ return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX);
+ case Intrinsic::atomic_swap:
+ return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP);
}
}
Chains[i] = Part.getValue(0);
}
- if (NumRegs == 1)
- Chain = Chains[0];
+ if (NumRegs == 1 || Flag)
+ // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
+ // flagged to it. That is the CopyToReg nodes and the user are considered
+ // a single scheduling unit. If we create a TokenFactor and return it as
+ // chain, then the TokenFactor is both a predecessor (operand) of the
+ // user as well as a successor (the TF operands are flagged to the user).
+ // c1, f1 = CopyToReg
+ // c2, f2 = CopyToReg
+ // c3 = TokenFactor c1, c2
+ // ...
+ // = op c3, ..., f2
+ Chain = Chains[NumRegs-1];
else
Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Chains[0], NumRegs);
}
}
// Otherwise, we couldn't allocate enough registers for this.
- return;
}
// Loop over all of the inputs, copying the operand values into the
// appropriate registers and processing the output regs.
RegsForValue RetValRegs;
-
+
// IndirectStoresToEmit - The set of stores to emit after the inline asm node.
std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
exit(1);
}
- if (!OpInfo.isIndirect) {
- // This is the result value of the call.
- assert(RetValRegs.Regs.empty() &&
- "Cannot have multiple output constraints yet!");
- assert(CS.getType() != Type::VoidTy && "Bad inline asm!");
- RetValRegs = OpInfo.AssignedRegs;
- } else {
+ // If this is an indirect operand, store through the pointer after the
+ // asm.
+ if (OpInfo.isIndirect) {
IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
OpInfo.CallOperandVal));
+ } else {
+ // This is the result value of the call.
+ assert(CS.getType() != Type::VoidTy && "Bad inline asm!");
+ // Concatenate this output onto the outputs list.
+ RetValRegs.append(OpInfo.AssignedRegs);
}
// Add information to the INLINEASM node to know that this register is
// Add NumOps>>3 registers to MatchedRegs.
RegsForValue MatchedRegs;
MatchedRegs.TLI = &TLI;
- MatchedRegs.ValueVTs.resize(1, InOperandVal.getValueType());
- MatchedRegs.RegVTs.resize(1, AsmNodeOperands[CurOp+1].getValueType());
+ MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
+ MatchedRegs.RegVTs.push_back(AsmNodeOperands[CurOp+1].getValueType());
for (unsigned i = 0, e = NumOps>>3; i != e; ++i) {
unsigned Reg =
cast<RegisterSDNode>(AsmNodeOperands[++CurOp])->getReg();
// and set it as the value of the call.
if (!RetValRegs.Regs.empty()) {
SDOperand Val = RetValRegs.getCopyFromRegs(DAG, Chain, &Flag);
-
- // If the result of the inline asm is a vector, it may have the wrong
- // width/num elts. Make sure to convert it to the right type with
+
+ // If any of the results of the inline asm is a vector, it may have the
+ // wrong width/num elts. This can happen for register classes that can
+ // contain multiple different value types. The preg or vreg allocated may
+ // not have the same VT as was expected. Convert it to the right type with
// bit_convert.
- if (MVT::isVector(Val.getValueType())) {
- const VectorType *VTy = cast<VectorType>(CS.getType());
- MVT::ValueType DesiredVT = TLI.getValueType(VTy);
-
- Val = DAG.getNode(ISD::BIT_CONVERT, DesiredVT, Val);
+ if (const StructType *ResSTy = dyn_cast<StructType>(CS.getType())) {
+ for (unsigned i = 0, e = ResSTy->getNumElements(); i != e; ++i) {
+ if (MVT::isVector(Val.Val->getValueType(i)))
+ Val = DAG.getNode(ISD::BIT_CONVERT,
+ TLI.getValueType(ResSTy->getElementType(i)), Val);
+ }
+ } else {
+ if (MVT::isVector(Val.getValueType()))
+ Val = DAG.getNode(ISD::BIT_CONVERT, TLI.getValueType(CS.getType()),
+ Val);
}
-
+
setValue(CS.getInstruction(), Val);
}
AU.setPreservesAll();
}
-
-
bool SelectionDAGISel::runOnFunction(Function &Fn) {
// Get alias analysis for load/store combining.
AA = &getAnalysis<AliasAnalysis>();
}
}
+/// IsFixedFrameObjectWithPosOffset - Check if object is a fixed frame object and
+/// whether object offset >= 0.
+static bool
+IsFixedFrameObjectWithPosOffset(MachineFrameInfo * MFI, SDOperand Op) {
+ if (!isa<FrameIndexSDNode>(Op)) return false;
+
+ FrameIndexSDNode * FrameIdxNode = dyn_cast<FrameIndexSDNode>(Op);
+ int FrameIdx = FrameIdxNode->getIndex();
+ return MFI->isFixedObjectIndex(FrameIdx) &&
+ MFI->getObjectOffset(FrameIdx) >= 0;
+}
+
+/// IsPossiblyOverwrittenArgumentOfTailCall - Check if the operand could
+/// possibly be overwritten when lowering the outgoing arguments in a tail
+/// call. Currently the implementation of this call is very conservative and
+/// assumes all arguments sourcing from FORMAL_ARGUMENTS or a CopyFromReg with
+/// virtual registers would be overwritten by direct lowering.
+static bool IsPossiblyOverwrittenArgumentOfTailCall(SDOperand Op,
+ MachineFrameInfo * MFI) {
+ RegisterSDNode * OpReg = NULL;
+ if (Op.getOpcode() == ISD::FORMAL_ARGUMENTS ||
+ (Op.getOpcode()== ISD::CopyFromReg &&
+ (OpReg = dyn_cast<RegisterSDNode>(Op.getOperand(1))) &&
+ (OpReg->getReg() >= TargetRegisterInfo::FirstVirtualRegister)) ||
+ (Op.getOpcode() == ISD::LOAD &&
+ IsFixedFrameObjectWithPosOffset(MFI, Op.getOperand(1))) ||
+ (Op.getOpcode() == ISD::MERGE_VALUES &&
+ Op.getOperand(Op.ResNo).getOpcode() == ISD::LOAD &&
+ IsFixedFrameObjectWithPosOffset(MFI, Op.getOperand(Op.ResNo).
+ getOperand(1))))
+ return true;
+ return false;
+}
+
/// CheckDAGForTailCallsAndFixThem - This Function looks for CALL nodes in the
/// DAG and fixes their tailcall attribute operand.
static void CheckDAGForTailCallsAndFixThem(SelectionDAG &DAG,
// eligible (no RET or the target rejects) the attribute is fixed to
// false. The TargetLowering::IsEligibleForTailCallOptimization function
// must correctly identify tail call optimizable calls.
- if (isMarkedTailCall &&
- (Ret==NULL ||
- !TLI.IsEligibleForTailCallOptimization(OpCall, OpRet, DAG))) {
+ if (!isMarkedTailCall) continue;
+ if (Ret==NULL ||
+ !TLI.IsEligibleForTailCallOptimization(OpCall, OpRet, DAG)) {
+ // Not eligible. Mark CALL node as non tail call.
SmallVector<SDOperand, 32> Ops;
unsigned idx=0;
- for(SDNode::op_iterator I =OpCall.Val->op_begin(),
- E=OpCall.Val->op_end(); I!=E; I++, idx++) {
+ for(SDNode::op_iterator I =OpCall.Val->op_begin(),
+ E = OpCall.Val->op_end(); I != E; I++, idx++) {
if (idx!=3)
Ops.push_back(*I);
- else
+ else
Ops.push_back(DAG.getConstant(false, TLI.getPointerTy()));
}
DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
+ } else {
+ // Look for tail call clobbered arguments. Emit a series of
+ // copyto/copyfrom virtual register nodes to protect them.
+ SmallVector<SDOperand, 32> Ops;
+ SDOperand Chain = OpCall.getOperand(0), InFlag;
+ unsigned idx=0;
+ for(SDNode::op_iterator I = OpCall.Val->op_begin(),
+ E = OpCall.Val->op_end(); I != E; I++, idx++) {
+ SDOperand Arg = *I;
+ if (idx > 4 && (idx % 2)) {
+ bool isByVal = cast<ARG_FLAGSSDNode>(OpCall.getOperand(idx+1))->
+ getArgFlags().isByVal();
+ MachineFunction &MF = DAG.getMachineFunction();
+ MachineFrameInfo *MFI = MF.getFrameInfo();
+ if (!isByVal &&
+ IsPossiblyOverwrittenArgumentOfTailCall(Arg, MFI)) {
+ MVT::ValueType VT = Arg.getValueType();
+ unsigned VReg = MF.getRegInfo().
+ createVirtualRegister(TLI.getRegClassFor(VT));
+ Chain = DAG.getCopyToReg(Chain, VReg, Arg, InFlag);
+ InFlag = Chain.getValue(1);
+ Arg = DAG.getCopyFromReg(Chain, VReg, VT, InFlag);
+ Chain = Arg.getValue(1);
+ InFlag = Arg.getValue(2);
+ }
+ }
+ Ops.push_back(Arg);
+ }
+ // Link in chain of CopyTo/CopyFromReg.
+ Ops[0] = Chain;
+ DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
}
}
}
projects / oota-llvm.git / blobdiff