First chunk of actually generating vector code for packed types. These

changes allow us to generate the following code:

_foo:
        li r2, 0
        lvx v0, r2, r3
        vaddfp v0, v0, v0
        stvx v0, r2, r3
        blr

for this llvm:

void %foo(<4 x float>* %a) {
entry:
        %tmp1 = load <4 x float>* %a
        %tmp2 = add <4 x float> %tmp1, %tmp1
        store <4 x float> %tmp2, <4 x float>* %a
        ret void
}


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24534 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 80 insertions, 17 deletions

diff --git a/include/llvm/CodeGen/ValueTypes.h b/include/llvm/CodeGen/ValueTypes.h
index 783c0f8acf..5a36a3eeec 100644
--- a/include/llvm/CodeGen/ValueTypes.h
+++ b/include/llvm/CodeGen/ValueTypes.h
@@ -45,10 +45,8 @@ namespace MVT {  // MVT = Machine Value Types
     isVoid         =  12,   // This has no value
     
     Vector         =  13,   // This is an abstract vector type, which will
-                            // be refined into a target vector type, or
-                            // scalarized.
-
-                            // These are 128 bit vectors of varying packed types
+                            // be expanded into a target vector type, or scalars
+                            // if no matching vector type is available.
     v16i8          =  14,   // 16 x i8
     v8i16          =  15,   //  8 x i16
     v4i32          =  16,   //  4 x i32
@@ -70,6 +68,21 @@ namespace MVT {  // MVT = Machine Value Types
     return (VT >= v16i8 && VT <= v2f64);
   }
 
+  /// getVectorType - Returns the ValueType that represents a vector NumElements
+  /// in length, where each element is of type VT.  If there is no ValueType
+  /// that represents this vector, a ValueType of Other is returned.
+  ///
+  static inline ValueType getVectorType(ValueType VT, unsigned NumElements) {
+    switch (VT) {
+    default: 
+      break;
+    case MVT::f32:
+      if (NumElements == 4) return MVT::v4f32;
+      break;
+    }
+    return MVT::Other;
+  }
+  
   static inline unsigned getSizeInBits(ValueType VT) {
     switch (VT) {
     default: assert(0 && "ValueType has no known size!");
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
index aad1708187..30d94337e9 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
@@ -520,12 +520,19 @@ void SelectionDAGLowering::visitBinary(User &I, unsigned IntOp, unsigned FPOp,
     const PackedType *PTy = cast<PackedType>(Ty);
     unsigned NumElements = PTy->getNumElements();
     MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
+    MVT::ValueType TVT = MVT::getVectorType(PVT, NumElements);
     
     // Immediately scalarize packed types containing only one element, so that
-    // the Legalize pass does not have to deal with them.
+    // the Legalize pass does not have to deal with them.  Similarly, if the
+    // abstract vector is going to turn into one that the target natively
+    // supports, generate that type now so that Legalize doesn't have to deal
+    // with that either.  These steps ensure that Legalize only has to handle
+    // vector types in its Expand case.
+    unsigned Opc = MVT::isFloatingPoint(PVT) ? FPOp : IntOp;
     if (NumElements == 1) {
-      unsigned Opc = MVT::isFloatingPoint(PVT) ? FPOp : IntOp;
       setValue(&I, DAG.getNode(Opc, PVT, Op1, Op2));
+    } else if (TVT != MVT::Other && TLI.isTypeLegal(TVT)) {
+      setValue(&I, DAG.getNode(Opc, TVT, Op1, Op2));
     } else {
       SDOperand Num = DAG.getConstant(NumElements, MVT::i32);
       SDOperand Typ = DAG.getValueType(PVT);
@@ -777,11 +784,14 @@ void SelectionDAGLowering::visitLoad(LoadInst &I) {
     const PackedType *PTy = cast<PackedType>(Ty);
     unsigned NumElements = PTy->getNumElements();
     MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
+    MVT::ValueType TVT = MVT::getVectorType(PVT, NumElements);
     
     // Immediately scalarize packed types containing only one element, so that
     // the Legalize pass does not have to deal with them.
     if (NumElements == 1) {
       L = DAG.getLoad(PVT, Root, Ptr, DAG.getSrcValue(I.getOperand(0)));
+    } else if (TVT != MVT::Other && TLI.isTypeLegal(TVT)) {
+      L = DAG.getLoad(TVT, Root, Ptr, DAG.getSrcValue(I.getOperand(0)));
     } else {
       L = DAG.getVecLoad(NumElements, PVT, Root, Ptr, 
                          DAG.getSrcValue(I.getOperand(0)));
diff --git a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
index 549f4539de..5b6032fd5a 100644
--- a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
+++ b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
@@ -73,6 +73,11 @@ namespace {
     /// load/store instruction, and return true if it should be an indexed [r+r]
     /// operation.
     bool SelectAddr(SDOperand Addr, SDOperand &Op1, SDOperand &Op2);
+    
+    /// SelectAddrIndexed - Given the specified addressed, force it to be
+    /// represented as an indexed [r+r] operation, rather than possibly
+    /// returning [r+imm] as SelectAddr may.
+    void SelectAddrIndexed(SDOperand Addr, SDOperand &Op1, SDOperand &Op2);
 
     SDOperand BuildSDIVSequence(SDNode *N);
     SDOperand BuildUDIVSequence(SDNode *N);
@@ -428,7 +433,7 @@ bool PPCDAGToDAGISel::SelectAddr(SDOperand Addr, SDOperand &Op1,
     }
   }
 
- if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Addr)) {
+  if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Addr)) {
     Op1 = getI32Imm(0);
     Op2 = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
     return false;
@@ -445,6 +450,26 @@ bool PPCDAGToDAGISel::SelectAddr(SDOperand Addr, SDOperand &Op1,
   return false;
 }
 
+/// SelectAddrIndexed - Given the specified addressed, force it to be
+/// represented as an indexed [r+r] operation, rather than possibly
+/// returning [r+imm] as SelectAddr may.
+void PPCDAGToDAGISel::SelectAddrIndexed(SDOperand Addr, SDOperand &Op1, 
+                                        SDOperand &Op2) {
+  if (Addr.getOpcode() == ISD::ADD) {
+    Op1 = Select(Addr.getOperand(0));
+    Op2 = Select(Addr.getOperand(1));
+    return;
+  }
+  
+  if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Addr)) {
+    Op1 = CurDAG->getTargetNode(PPC::LI, MVT::i32, getI32Imm(0));
+    Op2 = CurDAG->getTargetFrameIndex(FI->getIndex(), MVT::i32);
+    return;
+  }
+  Op1 = CurDAG->getTargetNode(PPC::LI, MVT::i32, getI32Imm(0));
+  Op2 = Select(Addr);
+}
+
 /// SelectCC - Select a comparison of the specified values with the specified
 /// condition code, returning the CR# of the expression.
 SDOperand PPCDAGToDAGISel::SelectCC(SDOperand LHS, SDOperand RHS,
@@ -916,9 +941,8 @@ SDOperand PPCDAGToDAGISel::Select(SDOperand Op) {
       }
     }
     
-    CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FADD : PPC::FADDS, Ty,
-                         Select(N->getOperand(0)), Select(N->getOperand(1)));
-    return SDOperand(N, 0);
+    // Other cases are autogenerated.
+    break;
   }
   case ISD::FSUB: {
     MVT::ValueType Ty = N->getValueType(0);
@@ -942,10 +966,9 @@ SDOperand PPCDAGToDAGISel::Select(SDOperand Op) {
         return SDOperand(N, 0);
       }
     }
-    CurDAG->SelectNodeTo(N, Ty == MVT::f64 ? PPC::FSUB : PPC::FSUBS, Ty,
-                         Select(N->getOperand(0)),
-                         Select(N->getOperand(1)));
-    return SDOperand(N, 0);
+
+    // Other cases are autogenerated.
+    break;
   }
   case ISD::SDIV: {
     // FIXME: since this depends on the setting of the carry flag from the srawi
@@ -1074,10 +1097,17 @@ SDOperand PPCDAGToDAGISel::Select(SDOperand Op) {
   case ISD::ZEXTLOAD:
   case ISD::SEXTLOAD: {
     SDOperand Op1, Op2;
-    bool isIdx = SelectAddr(N->getOperand(1), Op1, Op2);
-
+    // If this is a vector load, then force this to be indexed addressing, since
+    // altivec does not have immediate offsets for loads.
+    bool isIdx = true;
+    if (N->getOpcode() == ISD::LOAD && MVT::isVector(N->getValueType(0))) { 
+      SelectAddrIndexed(N->getOperand(1), Op1, Op2);
+    } else {
+      isIdx = SelectAddr(N->getOperand(1), Op1, Op2);
+    }
     MVT::ValueType TypeBeingLoaded = (N->getOpcode() == ISD::LOAD) ?
       N->getValueType(0) : cast<VTSDNode>(N->getOperand(3))->getVT();
+
     unsigned Opc;
     switch (TypeBeingLoaded) {
     default: N->dump(); assert(0 && "Cannot load this type!");
@@ -1093,6 +1123,7 @@ SDOperand PPCDAGToDAGISel::Select(SDOperand Op) {
     case MVT::i32: Opc = isIdx ? PPC::LWZX : PPC::LWZ; break;
     case MVT::f32: Opc = isIdx ? PPC::LFSX : PPC::LFS; break;
     case MVT::f64: Opc = isIdx ? PPC::LFDX : PPC::LFD; break;
+    case MVT::v4f32: Opc = PPC::LVX; break;
     }
 
     // If this is an f32 -> f64 load, emit the f32 load, then use an 'extending
@@ -1119,7 +1150,15 @@ SDOperand PPCDAGToDAGISel::Select(SDOperand Op) {
   case ISD::TRUNCSTORE:
   case ISD::STORE: {
     SDOperand AddrOp1, AddrOp2;
-    bool isIdx = SelectAddr(N->getOperand(2), AddrOp1, AddrOp2);
+    // If this is a vector store, then force this to be indexed addressing,
+    // since altivec does not have immediate offsets for stores.
+    bool isIdx = true;
+    if (N->getOpcode() == ISD::STORE && 
+        MVT::isVector(N->getOperand(1).getValueType())) {
+      SelectAddrIndexed(N->getOperand(2), AddrOp1, AddrOp2);
+    } else {
+      isIdx = SelectAddr(N->getOperand(2), AddrOp1, AddrOp2);
+    }
 
     unsigned Opc;
     if (N->getOpcode() == ISD::STORE) {
@@ -1128,6 +1167,7 @@ SDOperand PPCDAGToDAGISel::Select(SDOperand Op) {
       case MVT::i32: Opc = isIdx ? PPC::STWX  : PPC::STW; break;
       case MVT::f64: Opc = isIdx ? PPC::STFDX : PPC::STFD; break;
       case MVT::f32: Opc = isIdx ? PPC::STFSX : PPC::STFS; break;
+      case MVT::v4f32: Opc = PPC::STVX;
       }
     } else { //ISD::TRUNCSTORE
       switch(cast<VTSDNode>(N->getOperand(4))->getVT()) {