Part 1 to fix x86_64 fp128 calling convention.

[chh]

Summary

This is part 1 subset of http://reviews.llvm.org/D11438,
without the .td and new fp128 unit tests.

Most of the changes here only apply to the new x86_64 f128 type configuration,
which is not enable in this part.

Bugs:

• https://llvm.org/bugs/show_bug.cgi?id=24109
• https://llvm.org/bugs/show_bug.cgi?id=23897

Changes:

• Relax type legalization checks to accept new f128 type configuration, whose TypeAction is TypeSoftenFloat, not TypeLegal, but also has TLI.isTypeLegal true.
• Relax GetSoftenedFloat to return in some cases f128 type SDValue, which is TLI.isTypeLegal but not "softened" to i128 node.
• Allow customized FABS, FNEG, FCOPYSIGN on new f128 type configuration, to generate optimized bitwise operators for libm functions. Enhance related Lower* functions to handle f128 type.
• Enhance DAGTypeLegalizer::run, SoftenFloatResult, and related functions to keep new f128 type in register, and convert f128 operators to library calls.
• Fix Combiner, Emitter, Legalizer routines that did not handle f128 type.
• Add ExpandConstant to handle i128 constants, ExpandNode to handle ISD::Constant node.
• Add one more parameter to getCommonSubClass and firstCommonClass, to guarantee that returned common sub class will contain the specified simple value type. This extra parameter is used by EmitCopyFromReg in InstrEmitter.cpp.
• Fix infinite loop in getTypeLegalizationCost when f128 is the value type.
• Fix printOperand to handle null operand.
• Enhance ISD::BITCAST node to handle f128 constant.
• Expand new f128 type for BR_CC, SELECT_CC, SELECT, SETCC nodes.
• Enhance X86AsmPrinter to emit f128 values in comments.

[davidxl]

lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp

52 ↗

(On Diff #41578)

Push this change down to cases where No softening is needed.

73 ↗

(On Diff #41578)

Unrelated format change.

117 ↗

(On Diff #41578)

Put those comments and replace code into a small wrapper -- 'replaceValueIfNeeded'.

734 ↗

(On Diff #41578)

This comment seems out of place. Either remove it or put it in better place.

741 ↗

(On Diff #41578)

Move this comment also into the wrapper function. Explain better why it does not need to be softened again -- e.g, ReplaceValueWith is called eagerly ...

742 ↗

(On Diff #41578)

Move this into a helper function with a proper name? Also add comment why only these subset of opcodes can be skipped.

785 ↗

(On Diff #41578)

can be held.

787 ↗

(On Diff #41578)

Is the first condition needed? If Res.getNode() != N, ReplaceValueWith should also be called (see below), so this condition does not look correct.

797 ↗

(On Diff #41578)

typo re-analyze

lib/CodeGen/SelectionDAG/LegalizeTypes.cpp

241 ↗

(On Diff #41578)

Move this comment to SoftFloatResult where short cut was taken:

' Return false to tell caller to scan operands ...'

lib/CodeGen/SelectionDAG/LegalizeTypes.h

391 ↗

(On Diff #41578)

In that case,

394 ↗

(On Diff #41578)

Put it in a helper and add some DEBUG check what kind of Opcode are expected to not 'softened' at this point.

[chh]

lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp

117 ↗

(On Diff #41578)

Use a new small wrapper called ReplaceSoftenFloatResult.

787 ↗

(On Diff #41578)

Yes, it is necessary. It's not about Res.getNode()!=N.
When Res.getNode() == N,
(a) if LegalInHWReg, return false and do not reanalyze
(b) Otherwise, at line 799, return true and reanalyze

[davidxl]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

8776 ↗

(On Diff #41671)

Move the comment and additional check into a small helper for readability:

if ( (.... || ...) && legalToSimplifyCopysign(N1))

lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp

122 ↗

(On Diff #41671)

the caller.

775 ↗

(On Diff #41671)

llvm_unreachable? When this function is called, isLegalInHWReg must be true.

lib/CodeGen/SelectionDAG/LegalizeTypes.h

400 ↗

(On Diff #41671)

if (!SoftenedOp.getNode()) {

assert(isSimpleLegalType(...));  // or should it be assert(isLegalInHWReg(...)) ?
return Op;

}

414 ↗

(On Diff #41671)

It is clearer to do:

if (!isLegal...)
    return;

if (...)

463 ↗

(On Diff #41671)

soften -> softened

464 ↗

(On Diff #41671)

because the operand has already been softened in SoftenFloatResult with calls to ReplaceValueWith.

[silvas]

Each of your bullet points under Changes: sounds like it is about the right size for individual patches. Can you split this up?

[chh]

Sean, only the "Fix printOperand to handle null operand" in SelectionDAGDumper.cpp might have some value by itself.
All other changes are not useful before x86_64 f128 type is configured to fix the calling convention problem.
This part 1 change should have no impact at all before the rest of http://reviews.llvm.org/D11438 is merged in.
Getting this part 1 in first is just to check for any regression.
I think all these changes should go in or be reverted when necessary.
It would be easier to keep track of these changes together, right?

[chh]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

8776 ↗

(On Diff #41671)

Maybe you want something like IsOkayToConvertRoundOrExtendToCopySign(N1)?
The name legalToSimplifyCopysign(N1) here seems confusing,
as there are other 6 cases before this one to combine FABS, FNEG, and FCOPYSIGN operands
to FCOPYSIGN or FABS, FNEG, etc.
Maybe we could keep this condition close to its single user for now
until we know a clear abstraction?

lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp

775 ↗

(On Diff #41671)

Yes, but it should be better for this function to return false,
then the current caller will reuse existing dbg dump and llvm_unreachable.

lib/CodeGen/SelectionDAG/LegalizeTypes.h

400 ↗

(On Diff #41671)

I am not sure about assert here.
Original code will assert(SoftenOp.getNode() && ...) after RemapValue(SoftenedOp);
I cannot prove any error there, so I only want to have a condition here to return Op.

414 ↗

(On Diff #41671)

Would it be acceptable to test (NewRes.getNode() != N) first, as it should be faster than isLegalInHWReg and not always true.

[davidxl]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

8776 ↗

(On Diff #41693)

The comment of the method should clarify the intention -- I am not worried about naming.

lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp

775 ↗

(On Diff #41693)

ok -- better change the name of the function to 'CanSkipSoft...'

lib/CodeGen/SelectionDAG/LegalizeTypes.h

400 ↗

(On Diff #41693)

With the original code, this branch should never be taken -- so the new assertion should be safe, right?

414 ↗

(On Diff #41693)

ok

[silvas]

In D15134#300978, @chh wrote:

It would be easier to keep track of these changes together, right?

No, it is actually the opposite. The more fine-grained the individual commits, the easier it is for a reviewer to review, and the easier it is to identify what caused an issue if an issue does show up. Tests are also easier to write for fine-grained commits (or, it is easier to identify that a fine-grained commit does not change externally visible behavior and so doesn't require a test).

Also, the changes you are proposing here seem to fix bugs / introduce new functionality. Therefore tests are required.

[chh]

Sean, could you review the test cases in http://reviews.llvm.org/D11438 too?

New features (or bug fixes) introduced in this patch really should only be
executed when x86_64 f128 type is configured to stay in SSE registers.
We don't turn on that configuration in this partial patch, so all unit
tests in D11438 won't apply. Is there some new test we can add to test
other types for this patch?

After my initial submit of D11438 in Jul 22 2015, I have received virtually
no comment on those unit tests yet. I do feel the risk of test coverage for
both new x86_64 f128 config and existing configurations.

I explained before in patch D11438 comments that a partial change of x86_64
f128 is worse than no change at all, as we would rather have a working
non-gcc compatible x86_64 long double type than a buggy type or crashing
compiler. So a partial commit like this has only benefit to verify partial
regression to non x86_64 f128 types, which has lower possibility and I have
no trouble to revert them all at once.

The main battle will be for Android platform to verify the correctness of
x86_64 f128 type after the configuration is submitted in the follow up
patch.

[chh]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

8776 ↗

(On Diff #41693)

Fine. Please see the new split out function CanCombineFCOPYSIGN_EXTEND_ROUND.

lib/CodeGen/SelectionDAG/LegalizeTypes.h

400 ↗

(On Diff #41693)

Right, but the new assertion will be different from the old assertion at line 403 when it happens. As we are so carefully to avoid change of behavior to non x86_64 f128 type, I would avoid the change to assertion too.

[davidxl]

f128 related changes are pretty isolated at this point, so I will lgtm and the further changes can iterated later.

[chh]

David, thanks a lot for the review. This patch won't be this clean without your suggestions.
I'm going to submit this and watch for any regression.

[spatel]

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

8728–8740

Why is an x86-specific limitation being handled in DAGCombiner? Maybe no target supports f128, so it doesn't matter, but this seems wrong.

[chh]

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

8728–8740

I thought this optimization won't work for f128 given that current
llvm targets I know do not keep f128 in one register and have not defined
required instruction selection patterns. Instead of keeping this
optimization here for all targets and disable it in a target specific
way, it seems simpler to disable it here for f128 and enable such optimization
for any target later.

That being said, I didn't check all ad-hoc llvm floating point optimizations.
If my assumption is wrong or if there is a simpler way to handle this,
please feel free to suggest any different implementation.
Thanks.

llvm/trunk/include/llvm/Target/TargetLowering.h

public:
   };
 
   /// This enum indicates whether a types are legal for a target, and if not,
   /// what action should be used to make them valid.
   enum LegalizeTypeAction {
     TypeLegal,           // The target natively supports this type.
     TypePromoteInteger,  // Replace this integer with a larger one.
     TypeExpandInteger,   // Split this integer into two of half the size.
-    TypeSoftenFloat,     // Convert this float to a same size integer type.
+    TypeSoftenFloat,     // Convert this float to a same size integer type,
+                         // if an operation is not supported in target HW.
     TypeExpandFloat,     // Split this float into two of half the size.
     TypeScalarizeVector, // Replace this one-element vector with its element.
     TypeSplitVector,     // Split this vector into two of half the size.
     TypeWidenVector,     // This vector should be widened into a larger vector.
     TypePromoteFloat     // Replace this float with a larger one.
   };
 
   /// LegalizeKind holds the legalization kind that needs to happen to EVT
private:
   /// For each condition code (ISD::CondCode) keep a LegalizeAction that
   /// indicates how instruction selection should deal with the condition code.
   ///
   /// Because each CC action takes up 4 bits, we need to have the array size be
   /// large enough to fit all of the value types. This can be done by rounding
   /// up the MVT::LAST_VALUETYPE value to the next multiple of 8.
   uint32_t CondCodeActions[ISD::SETCC_INVALID][(MVT::LAST_VALUETYPE + 7) / 8];
 
+protected:
   ValueTypeActionImpl ValueTypeActions;
 
 private:
   LegalizeKind getTypeConversion(LLVMContext &Context, EVT VT) const;
 
 private:
   std::vector<std::pair<MVT, const TargetRegisterClass*> > AvailableRegClasses;
 

llvm/trunk/include/llvm/Target/TargetRegisterInfo.h

public:
 
   /// Returns the name of the register class.
   const char *getRegClassName(const TargetRegisterClass *Class) const {
     return MCRegisterInfo::getRegClassName(Class->MC);
   }
 
   /// Find the largest common subclass of A and B.
   /// Return NULL if there is no common subclass.
+  /// The common subclass should contain
+  /// simple value type SVT if it is not the Any type.
   const TargetRegisterClass *
   getCommonSubClass(const TargetRegisterClass *A,
-                    const TargetRegisterClass *B) const;
+                    const TargetRegisterClass *B,
+                    const MVT::SimpleValueType SVT =
+                    MVT::SimpleValueType::Any) const;
 
   /// Returns a TargetRegisterClass used for pointer values.
   /// If a target supports multiple different pointer register classes,
   /// kind specifies which one is indicated.
   virtual const TargetRegisterClass *
   getPointerRegClass(const MachineFunction &MF, unsigned Kind=0) const {
     llvm_unreachable("Target didn't implement getPointerRegClass!");
   }

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

SDValue DAGCombiner::visitFSQRT(SDNode *N) {
   SDValue ZeroCmp = DAG.getSetCC(DL, CCVT, N->getOperand(0), Zero, ISD::SETEQ);
   AddToWorklist(ZeroCmp.getNode());
   AddToWorklist(RV.getNode());
 
   return DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT, DL, VT,
                      ZeroCmp, Zero, RV);
 }
 
+static inline bool CanCombineFCOPYSIGN_EXTEND_ROUND(SDNode *N) {
+  // copysign(x, fp_extend(y)) -> copysign(x, y)
+  // copysign(x, fp_round(y)) -> copysign(x, y)
+  // Do not optimize out type conversion of f128 type yet.
+  // For some target like x86_64, configuration is changed
+  // to keep one f128 value in one SSE register, but
+  // instruction selection cannot handle FCOPYSIGN on
+  // SSE registers yet.
+  SDValue N1 = N->getOperand(1);
+  EVT N1VT = N1->getValueType(0);
+  EVT N1Op0VT = N1->getOperand(0)->getValueType(0);
+  return (N1.getOpcode() == ISD::FP_EXTEND ||
+          N1.getOpcode() == ISD::FP_ROUND) &&
+         (N1VT == N1Op0VT || N1Op0VT != MVT::f128);
+}
+
 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
   ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
   ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
   EVT VT = N->getValueType(0);
 
   if (N0CFP && N1CFP)  // Constant fold
SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
 
   // copysign(x, copysign(y,z)) -> copysign(x, z)
   if (N1.getOpcode() == ISD::FCOPYSIGN)
     return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
                        N0, N1.getOperand(1));
 
   // copysign(x, fp_extend(y)) -> copysign(x, y)
   // copysign(x, fp_round(y)) -> copysign(x, y)
-  if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
+  if (CanCombineFCOPYSIGN_EXTEND_ROUND(N))
     return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
                        N0, N1.getOperand(0));
 
   return SDValue();
 }
 
 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
   SDValue N0 = N->getOperand(0);

llvm/trunk/lib/CodeGen/SelectionDAG/InstrEmitter.cpp

for (SDNode *User : Node->uses()) {
             if (i+II.getNumDefs() < II.getNumOperands()) {
               RC = TRI->getAllocatableClass(
                 TII->getRegClass(II, i+II.getNumDefs(), TRI, *MF));
             }
             if (!UseRC)
               UseRC = RC;
             else if (RC) {
               const TargetRegisterClass *ComRC =
-                TRI->getCommonSubClass(UseRC, RC);
+                TRI->getCommonSubClass(UseRC, RC, VT.SimpleTy);
               // If multiple uses expect disjoint register classes, we emit
               // copies in AddRegisterOperand.
               if (ComRC)
                 UseRC = ComRC;
             }
           }
         }
       }

llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

private:
   SDValue ExpandBSWAP(SDValue Op, SDLoc dl);
   SDValue ExpandBitCount(unsigned Opc, SDValue Op, SDLoc dl);
 
   SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
   SDValue ExpandInsertToVectorThroughStack(SDValue Op);
   SDValue ExpandVectorBuildThroughStack(SDNode* Node);
 
   SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
+  SDValue ExpandConstant(ConstantSDNode *CP);
 
   // if ExpandNode returns false, LegalizeOp falls back to ConvertNodeToLibcall
   bool ExpandNode(SDNode *Node);
   void ConvertNodeToLibcall(SDNode *Node);
   void PromoteNode(SDNode *Node);
 
 public:
   // Node replacement helpers
SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
   }
   SDValue Result =
       DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
                   MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
                   false, false, false, Alignment);
   return Result;
 }
 
+/// Expands the Constant node to a load from the constant pool.
+SDValue SelectionDAGLegalize::ExpandConstant(ConstantSDNode *CP) {
+  SDLoc dl(CP);
+  EVT VT = CP->getValueType(0);
+  SDValue CPIdx = DAG.getConstantPool(CP->getConstantIntValue(),
+                                      TLI.getPointerTy(DAG.getDataLayout()));
+  unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
+  SDValue Result =
+    DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
+                MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
+                false, false, false, Alignment);
+  return Result;
+}
+
 /// Expands an unaligned store to 2 half-size stores.
 static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
                                  const TargetLowering &TLI,
                                  SelectionDAGLegalize *DAGLegalize) {
   assert(ST->getAddressingMode() == ISD::UNINDEXED &&
          "unaligned indexed stores not implemented!");
   SDValue Chain = ST->getChain();
   SDValue Ptr = ST->getBasePtr();

 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
   DEBUG(dbgs() << "\nLegalizing: "; Node->dump(&DAG));
 
   if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
     return;
 
 #ifndef NDEBUG
   for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
-    assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
-             TargetLowering::TypeLegal &&
+    assert((TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
+              TargetLowering::TypeLegal ||
+            TLI.isTypeLegal(Node->getValueType(i))) &&
            "Unexpected illegal type!");
 
   for (const SDValue &Op : Node->op_values())
-    assert((TLI.getTypeAction(*DAG.getContext(),
-                              Op.getValueType()) == TargetLowering::TypeLegal ||
-                              Op.getOpcode() == ISD::TargetConstant) &&
-                              "Unexpected illegal type!");
+    assert((TLI.getTypeAction(*DAG.getContext(), Op.getValueType()) ==
+              TargetLowering::TypeLegal ||
+            TLI.isTypeLegal(Op.getValueType()) ||
 #endif
 
   // Figure out the correct action; the way to query this varies by opcode
   TargetLowering::LegalizeAction Action = TargetLowering::Legal;
   bool SimpleFinishLegalizing = true;
   switch (Node->getOpcode()) {
   case ISD::INTRINSIC_W_CHAIN:
   case ISD::INTRINSIC_WO_CHAIN:
bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
   case ISD::ConstantFP: {
     ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
     // Check to see if this FP immediate is already legal.
     // If this is a legal constant, turn it into a TargetConstantFP node.
     if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
       Results.push_back(ExpandConstantFP(CFP, true));
     break;
   }
+  case ISD::Constant: {
+    ConstantSDNode *CP = cast<ConstantSDNode>(Node);
+    Results.push_back(ExpandConstant(CP));
+    break;
+  }
   case ISD::FSUB: {
     EVT VT = Node->getValueType(0);
     if (TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
         TLI.isOperationLegalOrCustom(ISD::FNEG, VT)) {
       const SDNodeFlags *Flags = &cast<BinaryWithFlagsSDNode>(Node)->Flags;
       Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
       Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1, Flags);
       Results.push_back(Tmp1);

llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp

return
     VT == MVT::f64 ? Call_F64 :
     VT == MVT::f80 ? Call_F80 :
     VT == MVT::f128 ? Call_F128 :
     VT == MVT::ppcf128 ? Call_PPCF128 :
     RTLIB::UNKNOWN_LIBCALL;
 }
 
 //===----------------------------------------------------------------------===//
-//  Result Float to Integer Conversion.
+//  Convert Float Results to Integer for Non-HW-supported Operations.
 //===----------------------------------------------------------------------===//
 
-void DAGTypeLegalizer::SoftenFloatResult(SDNode *N, unsigned ResNo) {
+bool DAGTypeLegalizer::SoftenFloatResult(SDNode *N, unsigned ResNo) {
   DEBUG(dbgs() << "Soften float result " << ResNo << ": "; N->dump(&DAG);
         dbgs() << "\n");
   SDValue R = SDValue();
 
   switch (N->getOpcode()) {
   default:
 #ifndef NDEBUG
     dbgs() << "SoftenFloatResult #" << ResNo << ": ";
     N->dump(&DAG); dbgs() << "\n";
 #endif
     llvm_unreachable("Do not know how to soften the result of this operator!");
 
+    case ISD::Register:
+    case ISD::CopyFromReg:
+    case ISD::CopyToReg:
+      assert(isLegalInHWReg(N->getValueType(ResNo)) &&
+             "Unsupported SoftenFloatRes opcode!");
+      // Only when isLegalInHWReg, we can skip check of the operands.
+      R = SDValue(N, ResNo);
+      break;
     case ISD::MERGE_VALUES:R = SoftenFloatRes_MERGE_VALUES(N, ResNo); break;
-    case ISD::BITCAST:     R = SoftenFloatRes_BITCAST(N); break;
+    case ISD::BITCAST:     R = SoftenFloatRes_BITCAST(N, ResNo); break;
     case ISD::BUILD_PAIR:  R = SoftenFloatRes_BUILD_PAIR(N); break;
-    case ISD::ConstantFP:
-      R = SoftenFloatRes_ConstantFP(cast<ConstantFPSDNode>(N));
-      break;
+    case ISD::ConstantFP:  R = SoftenFloatRes_ConstantFP(N, ResNo); break;
     case ISD::EXTRACT_VECTOR_ELT:
       R = SoftenFloatRes_EXTRACT_VECTOR_ELT(N); break;
-    case ISD::FABS:        R = SoftenFloatRes_FABS(N); break;
+    case ISD::FABS:        R = SoftenFloatRes_FABS(N, ResNo); break;
     case ISD::FMINNUM:     R = SoftenFloatRes_FMINNUM(N); break;
     case ISD::FMAXNUM:     R = SoftenFloatRes_FMAXNUM(N); break;
     case ISD::FADD:        R = SoftenFloatRes_FADD(N); break;
     case ISD::FCEIL:       R = SoftenFloatRes_FCEIL(N); break;
-    case ISD::FCOPYSIGN:   R = SoftenFloatRes_FCOPYSIGN(N); break;
+    case ISD::FCOPYSIGN:   R = SoftenFloatRes_FCOPYSIGN(N, ResNo); break;
     case ISD::FCOS:        R = SoftenFloatRes_FCOS(N); break;
     case ISD::FDIV:        R = SoftenFloatRes_FDIV(N); break;
     case ISD::FEXP:        R = SoftenFloatRes_FEXP(N); break;
     case ISD::FEXP2:       R = SoftenFloatRes_FEXP2(N); break;
     case ISD::FFLOOR:      R = SoftenFloatRes_FFLOOR(N); break;
     case ISD::FLOG:        R = SoftenFloatRes_FLOG(N); break;
     case ISD::FLOG2:       R = SoftenFloatRes_FLOG2(N); break;
     case ISD::FLOG10:      R = SoftenFloatRes_FLOG10(N); break;
     case ISD::FMA:         R = SoftenFloatRes_FMA(N); break;
     case ISD::FMUL:        R = SoftenFloatRes_FMUL(N); break;
     case ISD::FNEARBYINT:  R = SoftenFloatRes_FNEARBYINT(N); break;
-    case ISD::FNEG:        R = SoftenFloatRes_FNEG(N); break;
+    case ISD::FNEG:        R = SoftenFloatRes_FNEG(N, ResNo); break;
     case ISD::FP_EXTEND:   R = SoftenFloatRes_FP_EXTEND(N); break;
     case ISD::FP_ROUND:    R = SoftenFloatRes_FP_ROUND(N); break;
     case ISD::FP16_TO_FP:  R = SoftenFloatRes_FP16_TO_FP(N); break;
     case ISD::FPOW:        R = SoftenFloatRes_FPOW(N); break;
     case ISD::FPOWI:       R = SoftenFloatRes_FPOWI(N); break;
     case ISD::FREM:        R = SoftenFloatRes_FREM(N); break;
     case ISD::FRINT:       R = SoftenFloatRes_FRINT(N); break;
     case ISD::FROUND:      R = SoftenFloatRes_FROUND(N); break;
     case ISD::FSIN:        R = SoftenFloatRes_FSIN(N); break;
     case ISD::FSQRT:       R = SoftenFloatRes_FSQRT(N); break;
     case ISD::FSUB:        R = SoftenFloatRes_FSUB(N); break;
     case ISD::FTRUNC:      R = SoftenFloatRes_FTRUNC(N); break;
-    case ISD::LOAD:        R = SoftenFloatRes_LOAD(N); break;
-    case ISD::SELECT:      R = SoftenFloatRes_SELECT(N); break;
-    case ISD::SELECT_CC:   R = SoftenFloatRes_SELECT_CC(N); break;
+    case ISD::LOAD:        R = SoftenFloatRes_LOAD(N, ResNo); break;
+    case ISD::SELECT:      R = SoftenFloatRes_SELECT(N, ResNo); break;
+    case ISD::SELECT_CC:   R = SoftenFloatRes_SELECT_CC(N, ResNo); break;
     case ISD::SINT_TO_FP:
     case ISD::UINT_TO_FP:  R = SoftenFloatRes_XINT_TO_FP(N); break;
     case ISD::UNDEF:       R = SoftenFloatRes_UNDEF(N); break;
     case ISD::VAARG:       R = SoftenFloatRes_VAARG(N); break;
   }
 
   // If R is null, the sub-method took care of registering the result.
-  if (R.getNode())
+  if (R.getNode()) {
     SetSoftenedFloat(SDValue(N, ResNo), R);
+    ReplaceSoftenFloatResult(N, ResNo, R);
+  }
+  // Return true only if the node is changed,
+  // assuming that the operands are also converted when necessary.
+  // Otherwise, return false to tell caller to scan operands.
+  return R.getNode() && R.getNode() != N;
 }
 
-SDValue DAGTypeLegalizer::SoftenFloatRes_BITCAST(SDNode *N) {
+SDValue DAGTypeLegalizer::SoftenFloatRes_BITCAST(SDNode *N, unsigned ResNo) {
+  if (isLegalInHWReg(N->getValueType(ResNo)))
+    return SDValue(N, ResNo);
   return BitConvertToInteger(N->getOperand(0));
 }
 
 SDValue DAGTypeLegalizer::SoftenFloatRes_MERGE_VALUES(SDNode *N,
                                                       unsigned ResNo) {
   SDValue Op = DisintegrateMERGE_VALUES(N, ResNo);
   return BitConvertToInteger(Op);
 }
 
 SDValue DAGTypeLegalizer::SoftenFloatRes_BUILD_PAIR(SDNode *N) {
   // Convert the inputs to integers, and build a new pair out of them.
   return DAG.getNode(ISD::BUILD_PAIR, SDLoc(N),
                      TLI.getTypeToTransformTo(*DAG.getContext(),
                                               N->getValueType(0)),
                      BitConvertToInteger(N->getOperand(0)),
                      BitConvertToInteger(N->getOperand(1)));
 }
 
-SDValue DAGTypeLegalizer::SoftenFloatRes_ConstantFP(ConstantFPSDNode *N) {
-  return DAG.getConstant(N->getValueAPF().bitcastToAPInt(), SDLoc(N),
+SDValue DAGTypeLegalizer::SoftenFloatRes_ConstantFP(SDNode *N, unsigned ResNo) {
+  // When LegalInHWReg, we can load better from the constant pool.
+  if (isLegalInHWReg(N->getValueType(ResNo)))
+    return SDValue(N, ResNo);
+  ConstantFPSDNode *CN = cast<ConstantFPSDNode>(N);
+  return DAG.getConstant(CN->getValueAPF().bitcastToAPInt(), SDLoc(CN),
                          TLI.getTypeToTransformTo(*DAG.getContext(),
-                                                  N->getValueType(0)));
+                                                  CN->getValueType(0)));
 }
 
 SDValue DAGTypeLegalizer::SoftenFloatRes_EXTRACT_VECTOR_ELT(SDNode *N) {
   SDValue NewOp = BitConvertVectorToIntegerVector(N->getOperand(0));
   return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N),
                      NewOp.getValueType().getVectorElementType(),
                      NewOp, N->getOperand(1));
 }
 
-SDValue DAGTypeLegalizer::SoftenFloatRes_FABS(SDNode *N) {
+SDValue DAGTypeLegalizer::SoftenFloatRes_FABS(SDNode *N, unsigned ResNo) {
+  // When LegalInHWReg, FABS can be implemented as native bitwise operations.
+  if (isLegalInHWReg(N->getValueType(ResNo)))
+    return SDValue(N, ResNo);
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
   unsigned Size = NVT.getSizeInBits();
 
   // Mask = ~(1 << (Size-1))
   APInt API = APInt::getAllOnesValue(Size);
   API.clearBit(Size - 1);
   SDValue Mask = DAG.getConstant(API, SDLoc(N), NVT);
   SDValue Op = GetSoftenedFloat(N->getOperand(0));
return TLI.makeLibCall(DAG, GetFPLibCall(N->getValueType(0),
                                            RTLIB::CEIL_F32,
                                            RTLIB::CEIL_F64,
                                            RTLIB::CEIL_F80,
                                            RTLIB::CEIL_F128,
                                            RTLIB::CEIL_PPCF128),
                          NVT, Op, false, SDLoc(N)).first;
 }
 
-SDValue DAGTypeLegalizer::SoftenFloatRes_FCOPYSIGN(SDNode *N) {
+SDValue DAGTypeLegalizer::SoftenFloatRes_FCOPYSIGN(SDNode *N, unsigned ResNo) {
+  // When LegalInHWReg, FCOPYSIGN can be implemented as native bitwise operations.
+  if (isLegalInHWReg(N->getValueType(ResNo)))
+    return SDValue(N, ResNo);
   SDValue LHS = GetSoftenedFloat(N->getOperand(0));
   SDValue RHS = BitConvertToInteger(N->getOperand(1));
   SDLoc dl(N);
 
   EVT LVT = LHS.getValueType();
   EVT RVT = RHS.getValueType();
 
   unsigned LSize = LVT.getSizeInBits();
return TLI.makeLibCall(DAG, GetFPLibCall(N->getValueType(0),
                                            RTLIB::NEARBYINT_F32,
                                            RTLIB::NEARBYINT_F64,
                                            RTLIB::NEARBYINT_F80,
                                            RTLIB::NEARBYINT_F128,
                                            RTLIB::NEARBYINT_PPCF128),
                          NVT, Op, false, SDLoc(N)).first;
 }
 
-SDValue DAGTypeLegalizer::SoftenFloatRes_FNEG(SDNode *N) {
+SDValue DAGTypeLegalizer::SoftenFloatRes_FNEG(SDNode *N, unsigned ResNo) {
+  // When LegalInHWReg, FNEG can be implemented as native bitwise operations.
+  if (isLegalInHWReg(N->getValueType(ResNo)))
+    return SDValue(N, ResNo);
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
   SDLoc dl(N);
   // Expand Y = FNEG(X) -> Y = SUB -0.0, X
   SDValue Ops[2] = { DAG.getConstantFP(-0.0, dl, N->getValueType(0)),
                      GetSoftenedFloat(N->getOperand(0)) };
   return TLI.makeLibCall(DAG, GetFPLibCall(N->getValueType(0),
                                            RTLIB::SUB_F32,
                                            RTLIB::SUB_F64,
return TLI.makeLibCall(DAG, GetFPLibCall(N->getValueType(0),
                                            RTLIB::TRUNC_F32,
                                            RTLIB::TRUNC_F64,
                                            RTLIB::TRUNC_F80,
                                            RTLIB::TRUNC_F128,
                                            RTLIB::TRUNC_PPCF128),
                          NVT, Op, false, SDLoc(N)).first;
 }
 
-SDValue DAGTypeLegalizer::SoftenFloatRes_LOAD(SDNode *N) {
+SDValue DAGTypeLegalizer::SoftenFloatRes_LOAD(SDNode *N, unsigned ResNo) {
+  bool LegalInHWReg = isLegalInHWReg(N->getValueType(ResNo));
   LoadSDNode *L = cast<LoadSDNode>(N);
   EVT VT = N->getValueType(0);
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
   SDLoc dl(N);
 
   SDValue NewL;
   if (L->getExtensionType() == ISD::NON_EXTLOAD) {
     NewL = DAG.getLoad(L->getAddressingMode(), L->getExtensionType(),
                        NVT, dl, L->getChain(), L->getBasePtr(), L->getOffset(),
                        L->getPointerInfo(), NVT, L->isVolatile(),
                        L->isNonTemporal(), false, L->getAlignment(),
                        L->getAAInfo());
     // Legalized the chain result - switch anything that used the old chain to
     // use the new one.
+    if (N != NewL.getValue(1).getNode())
       ReplaceValueWith(SDValue(N, 1), NewL.getValue(1));
     return NewL;
   }
 
   // Do a non-extending load followed by FP_EXTEND.
   NewL = DAG.getLoad(L->getAddressingMode(), ISD::NON_EXTLOAD,
                      L->getMemoryVT(), dl, L->getChain(),
                      L->getBasePtr(), L->getOffset(), L->getPointerInfo(),
                      L->getMemoryVT(), L->isVolatile(),
                      L->isNonTemporal(), false, L->getAlignment(),
                      L->getAAInfo());
   // Legalized the chain result - switch anything that used the old chain to
   // use the new one.
   ReplaceValueWith(SDValue(N, 1), NewL.getValue(1));
-  return BitConvertToInteger(DAG.getNode(ISD::FP_EXTEND, dl, VT, NewL));
+  auto ExtendNode = DAG.getNode(ISD::FP_EXTEND, dl, VT, NewL);
+  if (LegalInHWReg)
+    return ExtendNode;
+  return BitConvertToInteger(ExtendNode);
 }
 
-SDValue DAGTypeLegalizer::SoftenFloatRes_SELECT(SDNode *N) {
+SDValue DAGTypeLegalizer::SoftenFloatRes_SELECT(SDNode *N, unsigned ResNo) {
+  if (isLegalInHWReg(N->getValueType(ResNo)))
+    return SDValue(N, ResNo);
   SDValue LHS = GetSoftenedFloat(N->getOperand(1));
   SDValue RHS = GetSoftenedFloat(N->getOperand(2));
   return DAG.getSelect(SDLoc(N),
                        LHS.getValueType(), N->getOperand(0), LHS, RHS);
 }
 
-SDValue DAGTypeLegalizer::SoftenFloatRes_SELECT_CC(SDNode *N) {
+SDValue DAGTypeLegalizer::SoftenFloatRes_SELECT_CC(SDNode *N, unsigned ResNo) {
+  if (isLegalInHWReg(N->getValueType(ResNo)))
+    return SDValue(N, ResNo);
   SDValue LHS = GetSoftenedFloat(N->getOperand(2));
   SDValue RHS = GetSoftenedFloat(N->getOperand(3));
   return DAG.getNode(ISD::SELECT_CC, SDLoc(N),
                      LHS.getValueType(), N->getOperand(0),
                      N->getOperand(1), LHS, RHS, N->getOperand(4));
 }
 
 SDValue DAGTypeLegalizer::SoftenFloatRes_UNDEF(SDNode *N) {
SDValue DAGTypeLegalizer::SoftenFloatRes_VAARG(SDNode *N) {
   SDLoc dl(N);
 
   SDValue NewVAARG;
   NewVAARG = DAG.getVAArg(NVT, dl, Chain, Ptr, N->getOperand(2),
                           N->getConstantOperandVal(3));
 
   // Legalized the chain result - switch anything that used the old chain to
   // use the new one.
+  if (N != NewVAARG.getValue(1).getNode())
     ReplaceValueWith(SDValue(N, 1), NewVAARG.getValue(1));
   return NewVAARG;
 }
 
 SDValue DAGTypeLegalizer::SoftenFloatRes_XINT_TO_FP(SDNode *N) {
   bool Signed = N->getOpcode() == ISD::SINT_TO_FP;
   EVT SVT = N->getOperand(0).getValueType();
   EVT RVT = N->getValueType(0);
   EVT NVT = EVT();
SDValue Op = DAG.getNode(Signed ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, dl,
                            NVT, N->getOperand(0));
   return TLI.makeLibCall(DAG, LC,
                          TLI.getTypeToTransformTo(*DAG.getContext(), RVT),
                          Op, Signed, dl).first;
 }
 
 
 //===----------------------------------------------------------------------===//
-//  Operand Float to Integer Conversion..
+//  Convert Float Operand to Integer for Non-HW-supported Operations.
 //===----------------------------------------------------------------------===//
 
 bool DAGTypeLegalizer::SoftenFloatOperand(SDNode *N, unsigned OpNo) {
   DEBUG(dbgs() << "Soften float operand " << OpNo << ": "; N->dump(&DAG);
         dbgs() << "\n");
   SDValue Res = SDValue();
 
   switch (N->getOpcode()) {
   default:
+    if (CanSkipSoftenFloatOperand(N, OpNo))
+      return false;
 #ifndef NDEBUG
     dbgs() << "SoftenFloatOperand Op #" << OpNo << ": ";
     N->dump(&DAG); dbgs() << "\n";
 #endif
     llvm_unreachable("Do not know how to soften this operator's operand!");
 
   case ISD::BITCAST:     Res = SoftenFloatOp_BITCAST(N); break;
   case ISD::BR_CC:       Res = SoftenFloatOp_BR_CC(N); break;
   case ISD::FP_EXTEND:   Res = SoftenFloatOp_FP_EXTEND(N); break;
   case ISD::FP_TO_FP16:  // Same as FP_ROUND for softening purposes
   case ISD::FP_ROUND:    Res = SoftenFloatOp_FP_ROUND(N); break;
   case ISD::FP_TO_SINT:  Res = SoftenFloatOp_FP_TO_SINT(N); break;
   case ISD::FP_TO_UINT:  Res = SoftenFloatOp_FP_TO_UINT(N); break;
   case ISD::SELECT_CC:   Res = SoftenFloatOp_SELECT_CC(N); break;
   case ISD::SETCC:       Res = SoftenFloatOp_SETCC(N); break;
-  case ISD::STORE:       Res = SoftenFloatOp_STORE(N, OpNo); break;
+  case ISD::STORE:
+    Res = SoftenFloatOp_STORE(N, OpNo);
+    // Do not try to analyze or soften this node again if the value is
+    // or can be held in a register. In that case, Res.getNode() should
+    // be equal to N.
+    if (Res.getNode() == N &&
+        isLegalInHWReg(N->getOperand(OpNo).getValueType()))
+      return false;
+    // Otherwise, we need to reanalyze and lower the new Res nodes.
+    break;
   }
 
   // If the result is null, the sub-method took care of registering results etc.
   if (!Res.getNode()) return false;
 
   // If the result is N, the sub-method updated N in place.  Tell the legalizer
-  // core about this.
+  // core about this to re-analyze.
   if (Res.getNode() == N)
     return true;
 
   assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 &&
          "Invalid operand expansion");
 
   ReplaceValueWith(SDValue(N, 0), Res);
   return false;
 }
 
+bool DAGTypeLegalizer::CanSkipSoftenFloatOperand(SDNode *N, unsigned OpNo) {
+  if (!isLegalInHWReg(N->getOperand(OpNo).getValueType()))
+    return false;
+  // When the operand type can be kept in registers, SoftenFloatResult
+  // will call ReplaceValueWith to replace all references and we can
+  // skip softening this operand.
+  switch (N->getOperand(OpNo).getOpcode()) {
+    case ISD::BITCAST:
+    case ISD::ConstantFP:
+    case ISD::CopyFromReg:
+    case ISD::CopyToReg:
+    case ISD::FABS:
+    case ISD::FCOPYSIGN:
+    case ISD::FNEG:
+    case ISD::Register:
+    case ISD::SELECT:
+    case ISD::SELECT_CC:
+      return true;
+  }
+  // For some opcodes, SoftenFloatResult handles all conversion of softening
+  // and replacing operands, so that there is no need to soften operands
+  // again, although such opcode could be scanned for other illegal operands.
+  switch (N->getOpcode()) {
+    case ISD::ConstantFP:
+    case ISD::CopyFromReg:
+    case ISD::CopyToReg:
+    case ISD::FABS:
+    case ISD::FCOPYSIGN:
+    case ISD::FNEG:
+    case ISD::Register:
+      return true;
+  }
+  return false;
+}
+
 SDValue DAGTypeLegalizer::SoftenFloatOp_BITCAST(SDNode *N) {
   return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0),
                      GetSoftenedFloat(N->getOperand(0)));
 }
 
 SDValue DAGTypeLegalizer::SoftenFloatOp_FP_EXTEND(SDNode *N) {
   // If we get here, the result must be legal but the source illegal.
   EVT SVT = N->getOperand(0).getValueType();

llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeTypes.h

TargetLowering::LegalizeTypeAction getTypeAction(EVT VT) const {
     return TLI.getTypeAction(*DAG.getContext(), VT);
   }
 
   /// isTypeLegal - Return true if this type is legal on this target.
   bool isTypeLegal(EVT VT) const {
     return TLI.getTypeAction(*DAG.getContext(), VT) == TargetLowering::TypeLegal;
   }
 
+  /// isSimpleLegalType - Return true if this is a simple legal type.
+  bool isSimpleLegalType(EVT VT) const {
+    return VT.isSimple() && TLI.isTypeLegal(VT);
+  }
+
+  /// isLegalInHWReg - Return true if this type can be passed in registers.
+  /// For example, x86_64's f128, should to be legally in registers
+  /// and only some operations converted to library calls or integer
+  /// bitwise operations.
+  bool isLegalInHWReg(EVT VT) const {
+    EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
+    return VT == NVT && isSimpleLegalType(VT);
+  }
+
   EVT getSetCCResultType(EVT VT) const {
     return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
   }
 
   /// IgnoreNodeResults - Pretend all of this node's results are legal.
   bool IgnoreNodeResults(SDNode *N) const {
     return N->getOpcode() == ISD::TargetConstant;
   }
private:
 
   void IntegerExpandSetCCOperands(SDValue &NewLHS, SDValue &NewRHS,
                                   ISD::CondCode &CCCode, SDLoc dl);
 
   //===--------------------------------------------------------------------===//
   // Float to Integer Conversion Support: LegalizeFloatTypes.cpp
   //===--------------------------------------------------------------------===//
 
-  /// GetSoftenedFloat - Given a processed operand Op which was converted to an
-  /// integer of the same size, this returns the integer.  The integer contains
-  /// exactly the same bits as Op - only the type changed.  For example, if Op
-  /// is an f32 which was softened to an i32, then this method returns an i32,
-  /// the bits of which coincide with those of Op.
+  /// GetSoftenedFloat - Given an operand Op of Float type, returns the integer
+  /// if the Op is not supported in target HW and converted to the integer.
+  /// The integer contains exactly the same bits as Op - only the type changed.
+  /// For example, if Op is an f32 which was softened to an i32, then this method
+  /// returns an i32, the bits of which coincide with those of Op.
+  /// If the Op can be efficiently supported in target HW or the operand must
+  /// stay in a register, the Op is not converted to an integer.
+  /// In that case, the given op is returned.
   SDValue GetSoftenedFloat(SDValue Op) {
     SDValue &SoftenedOp = SoftenedFloats[Op];
+    if (!SoftenedOp.getNode() &&
+        isSimpleLegalType(Op.getValueType()))
+      return Op;
     RemapValue(SoftenedOp);
     assert(SoftenedOp.getNode() && "Operand wasn't converted to integer?");
     return SoftenedOp;
   }
   void SetSoftenedFloat(SDValue Op, SDValue Result);
 
-  // Result Float to Integer Conversion.
-  void SoftenFloatResult(SDNode *N, unsigned OpNo);
+  // Call ReplaceValueWith(SDValue(N, ResNo), Res) if necessary.
+  void ReplaceSoftenFloatResult(SDNode *N, unsigned ResNo, SDValue &NewRes) {
+    // When the result type can be kept in HW registers, the converted
+    // NewRes node could have the same type. We can save the effort in
+    // cloning every user of N in SoftenFloatOperand or other legalization functions,
+    // by calling ReplaceValueWith here to update all users.
+    if (NewRes.getNode() != N && isLegalInHWReg(N->getValueType(ResNo)))
+      ReplaceValueWith(SDValue(N, ResNo), NewRes);
+  }
+
+  // Convert Float Results to Integer for Non-HW-supported Operations.
+  bool SoftenFloatResult(SDNode *N, unsigned ResNo);
   SDValue SoftenFloatRes_MERGE_VALUES(SDNode *N, unsigned ResNo);
-  SDValue SoftenFloatRes_BITCAST(SDNode *N);
+  SDValue SoftenFloatRes_BITCAST(SDNode *N, unsigned ResNo);
   SDValue SoftenFloatRes_BUILD_PAIR(SDNode *N);
-  SDValue SoftenFloatRes_ConstantFP(ConstantFPSDNode *N);
+  SDValue SoftenFloatRes_ConstantFP(SDNode *N, unsigned ResNo);
   SDValue SoftenFloatRes_EXTRACT_VECTOR_ELT(SDNode *N);
-  SDValue SoftenFloatRes_FABS(SDNode *N);
+  SDValue SoftenFloatRes_FABS(SDNode *N, unsigned ResNo);
   SDValue SoftenFloatRes_FMINNUM(SDNode *N);
   SDValue SoftenFloatRes_FMAXNUM(SDNode *N);
   SDValue SoftenFloatRes_FADD(SDNode *N);
   SDValue SoftenFloatRes_FCEIL(SDNode *N);
-  SDValue SoftenFloatRes_FCOPYSIGN(SDNode *N);
+  SDValue SoftenFloatRes_FCOPYSIGN(SDNode *N, unsigned ResNo);
   SDValue SoftenFloatRes_FCOS(SDNode *N);
   SDValue SoftenFloatRes_FDIV(SDNode *N);
   SDValue SoftenFloatRes_FEXP(SDNode *N);
   SDValue SoftenFloatRes_FEXP2(SDNode *N);
   SDValue SoftenFloatRes_FFLOOR(SDNode *N);
   SDValue SoftenFloatRes_FLOG(SDNode *N);
   SDValue SoftenFloatRes_FLOG2(SDNode *N);
   SDValue SoftenFloatRes_FLOG10(SDNode *N);
   SDValue SoftenFloatRes_FMA(SDNode *N);
   SDValue SoftenFloatRes_FMUL(SDNode *N);
   SDValue SoftenFloatRes_FNEARBYINT(SDNode *N);
-  SDValue SoftenFloatRes_FNEG(SDNode *N);
+  SDValue SoftenFloatRes_FNEG(SDNode *N, unsigned ResNo);
   SDValue SoftenFloatRes_FP_EXTEND(SDNode *N);
   SDValue SoftenFloatRes_FP16_TO_FP(SDNode *N);
   SDValue SoftenFloatRes_FP_ROUND(SDNode *N);
   SDValue SoftenFloatRes_FPOW(SDNode *N);
   SDValue SoftenFloatRes_FPOWI(SDNode *N);
   SDValue SoftenFloatRes_FREM(SDNode *N);
   SDValue SoftenFloatRes_FRINT(SDNode *N);
   SDValue SoftenFloatRes_FROUND(SDNode *N);
   SDValue SoftenFloatRes_FSIN(SDNode *N);
   SDValue SoftenFloatRes_FSQRT(SDNode *N);
   SDValue SoftenFloatRes_FSUB(SDNode *N);
   SDValue SoftenFloatRes_FTRUNC(SDNode *N);
-  SDValue SoftenFloatRes_LOAD(SDNode *N);
-  SDValue SoftenFloatRes_SELECT(SDNode *N);
-  SDValue SoftenFloatRes_SELECT_CC(SDNode *N);
+  SDValue SoftenFloatRes_LOAD(SDNode *N, unsigned ResNo);
+  SDValue SoftenFloatRes_SELECT(SDNode *N, unsigned ResNo);
+  SDValue SoftenFloatRes_SELECT_CC(SDNode *N, unsigned ResNo);
   SDValue SoftenFloatRes_UNDEF(SDNode *N);
   SDValue SoftenFloatRes_VAARG(SDNode *N);
   SDValue SoftenFloatRes_XINT_TO_FP(SDNode *N);
 
-  // Operand Float to Integer Conversion.
+  // Return true if we can skip softening the given operand or SDNode because
+  // it was soften before by SoftenFloatResult and references to the operand
+  // were replaced by ReplaceValueWith.
+  bool CanSkipSoftenFloatOperand(SDNode *N, unsigned OpNo);
+
+  // Convert Float Operand to Integer for Non-HW-supported Operations.
   bool SoftenFloatOperand(SDNode *N, unsigned OpNo);
   SDValue SoftenFloatOp_BITCAST(SDNode *N);
   SDValue SoftenFloatOp_BR_CC(SDNode *N);
   SDValue SoftenFloatOp_FP_EXTEND(SDNode *N);
   SDValue SoftenFloatOp_FP_ROUND(SDNode *N);
   SDValue SoftenFloatOp_FP_TO_SINT(SDNode *N);
   SDValue SoftenFloatOp_FP_TO_UINT(SDNode *N);
   SDValue SoftenFloatOp_SELECT_CC(SDNode *N);

llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeTypes.cpp

for (unsigned i = 0, NumResults = N->getNumValues(); i < NumResults; ++i) {
         PromoteIntegerResult(N, i);
         Changed = true;
         goto NodeDone;
       case TargetLowering::TypeExpandInteger:
         ExpandIntegerResult(N, i);
         Changed = true;
         goto NodeDone;
       case TargetLowering::TypeSoftenFloat:
-        SoftenFloatResult(N, i);
-        Changed = true;
+        Changed = SoftenFloatResult(N, i);
+        if (Changed)
           goto NodeDone;
+        // If not changed, the result type should be legally in register.
+        assert(isLegalInHWReg(ResultVT) &&
+               "Unchanged SoftenFloatResult should be legal in register!");
+        goto ScanOperands;
       case TargetLowering::TypeExpandFloat:
         ExpandFloatResult(N, i);
         Changed = true;
         goto NodeDone;
       case TargetLowering::TypeScalarizeVector:
         ScalarizeVectorResult(N, i);
         Changed = true;
         goto NodeDone;
#endif
 
   // In a debug build, scan all the nodes to make sure we found them all.  This
   // ensures that there are no cycles and that everything got processed.
 #ifndef NDEBUG
   for (SDNode &Node : DAG.allnodes()) {
     bool Failed = false;
 
     // Check that all result types are legal.
+    // A value type is illegal if its TypeAction is not TypeLegal,
+    // and TLI.RegClassForVT does not have a register class for this type.
+    // For example, the x86_64 target has f128 that is not TypeLegal,
+    // to have softened operators, but it also has FR128 register class to
+    // pass and return f128 values. Hence a legalized node can have f128 type.
     if (!IgnoreNodeResults(&Node))
       for (unsigned i = 0, NumVals = Node.getNumValues(); i < NumVals; ++i)
-        if (!isTypeLegal(Node.getValueType(i))) {
-          dbgs() << "Result type " << i << " illegal!\n";
+        if (!isTypeLegal(Node.getValueType(i)) &&
+            !TLI.isTypeLegal(Node.getValueType(i))) {
+          dbgs() << "Result type " << i << " illegal: ";
+          Node.dump();
           Failed = true;
         }
 
     // Check that all operand types are legal.
     for (unsigned i = 0, NumOps = Node.getNumOperands(); i < NumOps; ++i)
       if (!IgnoreNodeResults(Node.getOperand(i).getNode()) &&
-          !isTypeLegal(Node.getOperand(i).getValueType())) {
-        dbgs() << "Operand type " << i << " illegal!\n";
+          !isTypeLegal(Node.getOperand(i).getValueType()) &&
+          !TLI.isTypeLegal(Node.getOperand(i).getValueType())) {
+        dbgs() << "Operand type " << i << " illegal: ";
+        Node.getOperand(i).dump();
         Failed = true;
       }
 
     if (Node.getNodeId() != Processed) {
        if (Node.getNodeId() == NewNode)
          dbgs() << "New node not analyzed?\n";
        else if (Node.getNodeId() == Unanalyzed)
          dbgs() << "Unanalyzed node not noticed?\n";
void DAGTypeLegalizer::SetPromotedInteger(SDValue Op, SDValue Result) {
   AnalyzeNewValue(Result);
 
   SDValue &OpEntry = PromotedIntegers[Op];
   assert(!OpEntry.getNode() && "Node is already promoted!");
   OpEntry = Result;
 }
 
 void DAGTypeLegalizer::SetSoftenedFloat(SDValue Op, SDValue Result) {
-  assert(Result.getValueType() ==
-         TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType()) &&
+  // f128 of x86_64 could be kept in SSE registers,
+  // but sometimes softened to i128.
+  assert((Result.getValueType() ==
+          TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType()) ||
+          Op.getValueType() ==
+          TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType())) &&
          "Invalid type for softened float");
   AnalyzeNewValue(Result);
 
   SDValue &OpEntry = SoftenedFloats[Op];
-  assert(!OpEntry.getNode() && "Node is already converted to integer!");
+  // Allow repeated calls to save f128 type nodes
+  // or any node with type that transforms to itself.
+  // Many operations on these types are not softened.
+  assert((!OpEntry.getNode()||
+          Op.getValueType() ==
+          TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType())) &&
+         "Node is already converted to integer!");
   OpEntry = Result;
 }
 
 void DAGTypeLegalizer::SetPromotedFloat(SDValue Op, SDValue Result) {
   assert(Result.getValueType() ==
          TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType()) &&
          "Invalid type for promoted float");
   AnalyzeNewValue(Result);

llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp

void DAGTypeLegalizer::ExpandRes_BITCAST(SDNode *N, SDValue &Lo, SDValue &Hi) {
   // Handle some special cases efficiently.
   switch (getTypeAction(InVT)) {
     case TargetLowering::TypeLegal:
     case TargetLowering::TypePromoteInteger:
       break;
     case TargetLowering::TypePromoteFloat:
       llvm_unreachable("Bitcast of a promotion-needing float should never need"
                        "expansion");
-    case TargetLowering::TypeSoftenFloat:
-      // Convert the integer operand instead.
-      SplitInteger(GetSoftenedFloat(InOp), Lo, Hi);
+    case TargetLowering::TypeSoftenFloat: {
+      // Expand the floating point operand only if it was converted to integers.
+      // Otherwise, it is a legal type like f128 that can be saved in a register.
+      auto SoftenedOp = GetSoftenedFloat(InOp);
+      if (SoftenedOp == InOp)
+        break;
+      SplitInteger(SoftenedOp, Lo, Hi);
       Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
       Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
       return;
+    }
     case TargetLowering::TypeExpandInteger:
     case TargetLowering::TypeExpandFloat: {
       auto &DL = DAG.getDataLayout();
       // Convert the expanded pieces of the input.
       GetExpandedOp(InOp, Lo, Hi);
       if (TLI.hasBigEndianPartOrdering(InVT, DL) !=
           TLI.hasBigEndianPartOrdering(OutVT, DL))
         std::swap(Lo, Hi);

llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

case ISD::SINT_TO_FP: {
                                  APFloat::rmNearestTiesToEven);
       return getConstantFP(apf, DL, VT);
     }
     case ISD::BITCAST:
       if (VT == MVT::f16 && C->getValueType(0) == MVT::i16)
         return getConstantFP(APFloat(APFloat::IEEEhalf, Val), DL, VT);
       if (VT == MVT::f32 && C->getValueType(0) == MVT::i32)
         return getConstantFP(APFloat(APFloat::IEEEsingle, Val), DL, VT);
-      else if (VT == MVT::f64 && C->getValueType(0) == MVT::i64)
+      if (VT == MVT::f64 && C->getValueType(0) == MVT::i64)
         return getConstantFP(APFloat(APFloat::IEEEdouble, Val), DL, VT);
+      if (VT == MVT::f128 && C->getValueType(0) == MVT::i128)
+        return getConstantFP(APFloat(APFloat::IEEEquad, Val), DL, VT);
       break;
     case ISD::BSWAP:
       return getConstant(Val.byteSwap(), DL, VT, C->isTargetOpcode(),
                          C->isOpaque());
     case ISD::CTPOP:
       return getConstant(Val.countPopulation(), DL, VT, C->isTargetOpcode(),
                          C->isOpaque());
     case ISD::CTLZ:

llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp

void SDNode::printr(raw_ostream &OS, const SelectionDAG *G) const {
   OS << PrintNodeId(*this) << ": ";
   print_types(OS, G);
   OS << " = " << getOperationName(G);
   print_details(OS, G);
 }
 
 static bool printOperand(raw_ostream &OS, const SelectionDAG *G,
                          const SDValue Value) {
-  if (shouldPrintInline(*Value.getNode())) {
+  if (!Value.getNode()) {
+    OS << "<null>";
+    return false;
+  } else if (shouldPrintInline(*Value.getNode())) {
     OS << Value->getOperationName(G) << ':';
     Value->print_types(OS, G);
     Value->print_details(OS, G);
     return true;
   } else {
     OS << PrintNodeId(*Value.getNode());
     if (unsigned RN = Value.getResNo())
       OS << ':' << RN;

llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp

case ISD::BITCAST:
     // thing demanded, turn this into a FGETSIGN.
     if (!TLO.LegalOperations() &&
         !Op.getValueType().isVector() &&
         !Op.getOperand(0).getValueType().isVector() &&
         NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) &&
         Op.getOperand(0).getValueType().isFloatingPoint()) {
       bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType());
       bool i32Legal  = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32);
-      if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple()) {
+      if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple() &&
+           Op.getOperand(0).getValueType() != MVT::f128) {
+        // Cannot eliminate/lower SHL for f128 yet.
         EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32;
         // Make a FGETSIGN + SHL to move the sign bit into the appropriate
         // place.  We expect the SHL to be eliminated by other optimizations.
         SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0));
         unsigned OpVTSizeInBits = Op.getValueType().getSizeInBits();
         if (!OpVTLegal && OpVTSizeInBits > 32)
           Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign);
         unsigned ShVal = Op.getValueType().getSizeInBits()-1;

llvm/trunk/lib/CodeGen/TargetLoweringBase.cpp

while (true) {
     LegalizeKind LK = getTypeConversion(C, MTy);
 
     if (LK.first == TypeLegal)
       return std::make_pair(Cost, MTy.getSimpleVT());
 
     if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger)
       Cost *= 2;
 
+    // Do not loop with f128 type.
+    if (MTy == LK.second)
+      return std::make_pair(Cost, MTy.getSimpleVT());
+
     // Keep legalizing the type.
     MTy = LK.second;
   }
 }
 
 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
   if (!TM.getTargetTriple().isAndroid())
     return nullptr;

llvm/trunk/lib/CodeGen/TargetRegisterInfo.cpp

BitVector TargetRegisterInfo::getAllocatableSet(const MachineFunction &MF,
   Allocatable &= Reserved.flip();
 
   return Allocatable;
 }
 
 static inline
 const TargetRegisterClass *firstCommonClass(const uint32_t *A,
                                             const uint32_t *B,
-                                            const TargetRegisterInfo *TRI) {
+                                            const TargetRegisterInfo *TRI,
+                                            const MVT::SimpleValueType SVT =
+                                            MVT::SimpleValueType::Any) {
+  const MVT VT(SVT);
   for (unsigned I = 0, E = TRI->getNumRegClasses(); I < E; I += 32)
-    if (unsigned Common = *A++ & *B++)
-      return TRI->getRegClass(I + countTrailingZeros(Common));
+    if (unsigned Common = *A++ & *B++) {
+      const TargetRegisterClass *RC =
+          TRI->getRegClass(I + countTrailingZeros(Common));
+      if (SVT == MVT::SimpleValueType::Any || RC->hasType(VT))
+        return RC;
+    }
   return nullptr;
 }
 
 const TargetRegisterClass *
 TargetRegisterInfo::getCommonSubClass(const TargetRegisterClass *A,
-                                      const TargetRegisterClass *B) const {
+                                      const TargetRegisterClass *B,
+                                      const MVT::SimpleValueType SVT) const {
   // First take care of the trivial cases.
   if (A == B)
     return A;
   if (!A || !B)
     return nullptr;
 
   // Register classes are ordered topologically, so the largest common
   // sub-class it the common sub-class with the smallest ID.
-  return firstCommonClass(A->getSubClassMask(), B->getSubClassMask(), this);
+  return firstCommonClass(A->getSubClassMask(), B->getSubClassMask(), this, SVT);
 }
 
 const TargetRegisterClass *
 TargetRegisterInfo::getMatchingSuperRegClass(const TargetRegisterClass *A,
                                              const TargetRegisterClass *B,
                                              unsigned Idx) const {
   assert(A && B && "Missing register class");
   assert(Idx && "Bad sub-register index");

llvm/trunk/lib/Target/X86/X86MCInstLower.cpp

if (auto *C = getConstantFromPool(*MI, MI->getOperand(4))) {
         CS << "<";
         for (int i = 0, NumOperands = CV->getNumOperands(); i < NumOperands; ++i) {
           if (i != 0)
             CS << ",";
           Constant *COp = CV->getOperand(i);
           if (isa<UndefValue>(COp)) {
             CS << "u";
           } else if (auto *CI = dyn_cast<ConstantInt>(COp)) {
+            if (CI->getBitWidth() <= 64) {
               CS << CI->getZExtValue();
+            } else {
+              // print multi-word constant as (w0,w1)
+              auto Val = CI->getValue();
+              CS << "(";
+              for (int i = 0, N = Val.getNumWords(); i < N; ++i) {
+                if (i > 0)
+                  CS << ",";
+                CS << Val.getRawData()[i];
+              }
+              CS << ")";
+            }
           } else if (auto *CF = dyn_cast<ConstantFP>(COp)) {
             SmallString<32> Str;
             CF->getValueAPF().toString(Str);
             CS << Str;
           } else {
             CS << "?";
           }
         }