Remove all traces of v2[i,f]32 on SPU.

The "half vectors" are now widened to full size by the legalizer.
The only exception is in parameter passing, where half vectors are 
expanded. This causes changes to some dejagnu tests.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@111360 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 4 insertions, 110 deletions

diff --git a/lib/Target/CellSPU/SPUCallingConv.td b/lib/Target/CellSPU/SPUCallingConv.td
index dcd9a2e62e..04fa2ae866 100644
--- a/lib/Target/CellSPU/SPUCallingConv.td
+++ b/lib/Target/CellSPU/SPUCallingConv.td
@@ -21,8 +21,7 @@ class CCIfSubtarget<string F, CCAction A>
 
 // Return-value convention for Cell SPU: return value to be passed in reg 3-74
 def RetCC_SPU : CallingConv<[
-  CCIfType<[i8,i16,i32,i64,i128,f32,f64,v16i8,v8i16,v4i32,v2i64,v4f32,v2f64, 
-            v2i32, v2f32],
+  CCIfType<[i8,i16,i32,i64,i128,f32,f64,v16i8,v8i16,v4i32,v2i64,v4f32,v2f64],
   CCAssignToReg<[R3,   R4,  R5,  R6,  R7,  R8,  R9, R10, R11,
                  R12, R13, R14, R15, R16, R17, R18, R19, R20,
                  R21, R22, R23, R24, R25, R26, R27, R28, R29,
@@ -39,7 +38,7 @@ def RetCC_SPU : CallingConv<[
 //===----------------------------------------------------------------------===//
 def CCC_SPU : CallingConv<[
   CCIfType<[i8, i16, i32, i64, i128, f32, f64, 
-            v16i8, v8i16, v4i32, v4f32, v2i64, v2f64, v2i32, v2f32],
+            v16i8, v8i16, v4i32, v4f32, v2i64, v2f64],
             CCAssignToReg<[R3,   R4,  R5,  R6,  R7,  R8,  R9, R10, R11,
                            R12, R13, R14, R15, R16, R17, R18, R19, R20,
                            R21, R22, R23, R24, R25, R26, R27, R28, R29,
diff --git a/lib/Target/CellSPU/SPUISelLowering.cpp b/lib/Target/CellSPU/SPUISelLowering.cpp
index 1dc969b4b6..d5f3fec255 100644
--- a/lib/Target/CellSPU/SPUISelLowering.cpp
+++ b/lib/Target/CellSPU/SPUISelLowering.cpp
@@ -426,10 +426,6 @@ SPUTargetLowering::SPUTargetLowering(SPUTargetMachine &TM)
   addRegisterClass(MVT::v4f32, SPU::VECREGRegisterClass);
   addRegisterClass(MVT::v2f64, SPU::VECREGRegisterClass);
 
-  // "Odd size" vector classes that we're willing to support:
-  addRegisterClass(MVT::v2i32, SPU::VECREGRegisterClass);
-  addRegisterClass(MVT::v2f32, SPU::VECREGRegisterClass);
-
   for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
        i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
     MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
@@ -470,9 +466,6 @@ SPUTargetLowering::SPUTargetLowering(SPUTargetMachine &TM)
 
   setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
 
-  setOperationAction(ISD::STORE, MVT::v2i32, Custom);
-  setOperationAction(ISD::STORE, MVT::v2f32, Custom);
-
   setShiftAmountType(MVT::i32);
   setBooleanContents(ZeroOrNegativeOneBooleanContent);
 
@@ -1085,8 +1078,6 @@ SPUTargetLowering::LowerFormalArguments(SDValue Chain,
       case MVT::v4i32:
       case MVT::v8i16:
       case MVT::v16i8:
-      case MVT::v2i32:
-      case MVT::v2f32:
         ArgRegClass = &SPU::VECREGRegClass;
         break;
       }
@@ -1641,10 +1632,6 @@ LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
     SDValue T = DAG.getConstant(unsigned(SplatBits), VT.getVectorElementType());
     return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, T, T, T, T);
   }
-  case MVT::v2f32:
-  case MVT::v2i32: {
-    return SDValue();
-  }
   case MVT::v2i64: {
     return SPU::LowerV2I64Splat(VT, DAG, SplatBits, dl);
   }
@@ -1788,9 +1775,6 @@ static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
   } else if (EltVT == MVT::i16) {
     V2EltIdx0 = 8;
     maskVT = MVT::v8i16;
-  } else if (VecVT == MVT::v2i32 || VecVT == MVT::v2f32 ) {
-    V2EltIdx0 = 2;
-    maskVT = MVT::v4i32;
   } else if (EltVT == MVT::i32 || EltVT == MVT::f32) {
     V2EltIdx0 = 4;
     maskVT = MVT::v4i32;
@@ -1870,16 +1854,6 @@ static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
       for (unsigned j = 0; j < BytesPerElement; ++j)
         ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,MVT::i8));
     }
-    // For half vectors padd the mask with zeros for the second half.
-    // This is needed because mask is assumed to be full vector elsewhere in 
-    // the SPU backend. 
-    if(VecVT == MVT::v2i32 || VecVT == MVT::v2f32)
-    for( unsigned i = 0; i < 2; ++i )
-    {
-      for (unsigned j = 0; j < BytesPerElement; ++j)
-        ResultMask.push_back(DAG.getConstant(0,MVT::i8));
-    }
-
     SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
                                     &ResultMask[0], ResultMask.size());
     return DAG.getNode(SPUISD::SHUFB, dl, V1.getValueType(), V1, V2, VPermMask);
@@ -1909,7 +1883,6 @@ static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
     case MVT::v4f32: n_copies = 4; VT = MVT::f32; break;
     case MVT::v2i64: n_copies = 2; VT = MVT::i64; break;
     case MVT::v2f64: n_copies = 2; VT = MVT::f64; break;
-    case MVT::v2i32: n_copies = 2; VT = MVT::i32; break;
     }
 
     SDValue CValue = DAG.getConstant(CN->getZExtValue(), VT);
diff --git a/lib/Target/CellSPU/SPUInstrInfo.td b/lib/Target/CellSPU/SPUInstrInfo.td
index ff776fbe23..ca0fe00e37 100644
--- a/lib/Target/CellSPU/SPUInstrInfo.td
+++ b/lib/Target/CellSPU/SPUInstrInfo.td
@@ -62,9 +62,6 @@ let canFoldAsLoad = 1 in {
     def v4f32: LoadDFormVec<v4f32>;
     def v2f64: LoadDFormVec<v2f64>;
 
-    def v2i32: LoadDFormVec<v2i32>;
-    def v2f32: LoadDFormVec<v2f32>;
-
     def r128:  LoadDForm<GPRC>;
     def r64:   LoadDForm<R64C>;
     def r32:   LoadDForm<R32C>;
@@ -97,9 +94,6 @@ let canFoldAsLoad = 1 in {
     def v4f32: LoadAFormVec<v4f32>;
     def v2f64: LoadAFormVec<v2f64>;
 
-    def v2i32: LoadAFormVec<v2i32>;
-    def v2f32: LoadAFormVec<v2f32>;
-
     def r128:  LoadAForm<GPRC>;
     def r64:   LoadAForm<R64C>;
     def r32:   LoadAForm<R32C>;
@@ -132,9 +126,6 @@ let canFoldAsLoad = 1 in {
     def v4f32: LoadXFormVec<v4f32>;
     def v2f64: LoadXFormVec<v2f64>;
 
-    def v2i32: LoadXFormVec<v2i32>;
-    def v2f32: LoadXFormVec<v2f32>;
-
     def r128:  LoadXForm<GPRC>;
     def r64:   LoadXForm<R64C>;
     def r32:   LoadXForm<R32C>;
@@ -183,9 +174,6 @@ multiclass StoreDForms
   def v4f32: StoreDFormVec<v4f32>;
   def v2f64: StoreDFormVec<v2f64>;
 
-  def v2i32: StoreDFormVec<v2i32>;
-  def v2f32: StoreDFormVec<v2f32>;
-
   def r128:  StoreDForm<GPRC>;
   def r64:   StoreDForm<R64C>;
   def r32:   StoreDForm<R32C>;
@@ -216,9 +204,6 @@ multiclass StoreAForms
   def v4f32: StoreAFormVec<v4f32>;
   def v2f64: StoreAFormVec<v2f64>;
 
-  def v2i32: StoreAFormVec<v2i32>;
-  def v2f32: StoreAFormVec<v2f32>;
-
   def r128:  StoreAForm<GPRC>;
   def r64:   StoreAForm<R64C>;
   def r32:   StoreAForm<R32C>;
@@ -251,9 +236,6 @@ multiclass StoreXForms
   def v4f32: StoreXFormVec<v4f32>;
   def v2f64: StoreXFormVec<v2f64>;
 
-  def v2i32: StoreXFormVec<v2i32>;
-  def v2f32: StoreXFormVec<v2f32>;
-
   def r128:  StoreXForm<GPRC>;
   def r64:   StoreXForm<R64C>;
   def r32:   StoreXForm<R32C>;
@@ -613,7 +595,6 @@ class ARegInst<RegisterClass rclass>:
 multiclass AddInstruction {
   def v4i32: AVecInst<v4i32>;
   def v16i8: AVecInst<v16i8>;
-  def v2i32: AVecInst<v2i32>;
   def r32:   ARegInst<R32C>;
 }
 
@@ -678,11 +659,6 @@ def SFvec : RRForm<0b00000010000, (outs VECREG:$rT),
   "sf\t$rT, $rA, $rB", IntegerOp,
   [(set (v4i32 VECREG:$rT), (sub (v4i32 VECREG:$rB), (v4i32 VECREG:$rA)))]>;
 
-def SF2vec : RRForm<0b00000010000, (outs VECREG:$rT),
-                                  (ins VECREG:$rA, VECREG:$rB),
-  "sf\t$rT, $rA, $rB", IntegerOp,
-  [(set (v2i32 VECREG:$rT), (sub (v2i32 VECREG:$rB), (v2i32 VECREG:$rA)))]>;
-
 
 def SFr32 : RRForm<0b00000010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
   "sf\t$rT, $rA, $rB", IntegerOp,
@@ -841,10 +817,6 @@ def MPYUv4i32:
   MPYUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
            [/* no pattern */]>;
 
-def MPYUv2i32:
-  MPYUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
-           [/* no pattern */]>;
-
 def MPYUr16:
   MPYUInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB),
            [(set R32C:$rT, (mul (zext R16C:$rA), (zext R16C:$rB)))]>;
@@ -924,10 +896,6 @@ def MPYHv4i32:
     MPYHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
              [/* no pattern */]>;
 
-def MPYHv2i32:
-    MPYHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
-             [/* no pattern */]>;
-
 def MPYHr32:
     MPYHInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
              [/* no pattern */]>;
@@ -1517,13 +1485,6 @@ multiclass BitwiseOr
   def f32_v4f32: ORExtractElt<R32FP>;
   def f64_v2f64: ORExtractElt<R64FP>;
 
-  // half <-> full vector mappings
-  def v2i32_v4i32: ORCvtVecVec;
-  def v4i32_v2i32: ORCvtVecVec;
-  def v2f32_v4f32: ORCvtVecVec;
-  def v4f32_v2f32: ORCvtVecVec;
-
-
   // Conversion from vector to GPRC
   def i128_vec:  ORCvtVecGPRC;
 
@@ -1591,18 +1552,12 @@ def : Pat<(v8i16 (SPUprefslot2vec R16C:$rA)),
 def : Pat<(v4i32 (SPUprefslot2vec R32C:$rA)),
           (ORv4i32_i32 R32C:$rA)>;
 
-def : Pat<(v2i32 (SPUprefslot2vec R32C:$rA)),
-          (ORv4i32_i32 R32C:$rA)>;
-
 def : Pat<(v2i64 (SPUprefslot2vec R64C:$rA)),
           (ORv2i64_i64 R64C:$rA)>;
 
 def : Pat<(v4f32 (SPUprefslot2vec R32FP:$rA)),
           (ORv4f32_f32 R32FP:$rA)>;
 
-def : Pat<(v2f32 (SPUprefslot2vec R32FP:$rA)),
-          (ORv4f32_f32 R32FP:$rA)>;
-
 def : Pat<(v2f64 (SPUprefslot2vec R64FP:$rA)),
           (ORv2f64_f64 R64FP:$rA)>;
 
@@ -1618,33 +1573,15 @@ def : Pat<(SPUvec2prefslot (v8i16 VECREG:$rA)),
 def : Pat<(SPUvec2prefslot (v4i32 VECREG:$rA)),
           (ORi32_v4i32 VECREG:$rA)>;
 
-def : Pat<(SPUvec2prefslot (v2i32 VECREG:$rA)),
-          (ORi32_v4i32 VECREG:$rA)>;
-
 def : Pat<(SPUvec2prefslot (v2i64 VECREG:$rA)),
           (ORi64_v2i64 VECREG:$rA)>;
 
 def : Pat<(SPUvec2prefslot (v4f32 VECREG:$rA)),
           (ORf32_v4f32 VECREG:$rA)>;
 
-def : Pat<(SPUvec2prefslot (v2f32 VECREG:$rA)),
-          (ORf32_v4f32 VECREG:$rA)>;
-
 def : Pat<(SPUvec2prefslot (v2f64 VECREG:$rA)),
           (ORf64_v2f64 VECREG:$rA)>;
 
-// Conversions between 64 bit and 128 bit vectors. 
-
-def : Pat<(v4i32 (SPUhalf2vec (v2i32 VECREG:$rA))),
-          (ORv4i32_v2i32 (v2i32 VECREG:$rA))>;
-def : Pat<(v4f32 (SPUhalf2vec (v2f32 VECREG:$rA))),
-          (ORv4f32_v2f32 (v2f32 VECREG:$rA))>;
-
-def : Pat<(v2i32 (SPUvec2half (v4i32 VECREG:$rA))),
-          (ORv2i32_v4i32 VECREG:$rA)>;
-def : Pat<(v2f32 (SPUvec2half (v4f32 VECREG:$rA))),
-          (ORv2f32_v4f32 VECREG:$rA)>;
-
 // Load Register: This is an assembler alias for a bitwise OR of a register
 // against itself. It's here because it brings some clarity to assembly
 // language output.
@@ -2177,15 +2114,11 @@ multiclass ShuffleBytes
   def v8i16_m32 : SHUFBVecInst<v8i16, v4i32>;
   def v4i32     : SHUFBVecInst<v4i32, v16i8>;
   def v4i32_m32 : SHUFBVecInst<v4i32, v4i32>;
-  def v2i32     : SHUFBVecInst<v2i32, v16i8>;
-  def v2i32_m32 : SHUFBVecInst<v2i32, v4i32>;
   def v2i64     : SHUFBVecInst<v2i64, v16i8>;
   def v2i64_m32 : SHUFBVecInst<v2i64, v4i32>;
 
   def v4f32     : SHUFBVecInst<v4f32, v16i8>;
   def v4f32_m32 : SHUFBVecInst<v4f32, v4i32>;
-  def v2f32     : SHUFBVecInst<v2f32, v16i8>;
-  def v2f32_m32 : SHUFBVecInst<v2f32, v4i32>;
 
   def v2f64     : SHUFBVecInst<v2f64, v16i8>;
   def v2f64_m32 : SHUFBVecInst<v2f64, v4i32>;
@@ -3928,7 +3861,6 @@ class FAVecInst<ValueType vectype>:
 multiclass SFPAdd
 {
   def v4f32: FAVecInst<v4f32>;
-  def v2f32: FAVecInst<v2f32>;
   def f32:   FAInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
                     [(set R32FP:$rT, (fadd R32FP:$rA, R32FP:$rB))]>;
 }
@@ -3947,7 +3879,6 @@ class FSVecInst<ValueType vectype>:
 multiclass SFPSub
 {
   def v4f32: FSVecInst<v4f32>;
-  def v2f32: FSVecInst<v2f32>;
   def f32:   FSInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
                     [(set R32FP:$rT, (fsub R32FP:$rA, R32FP:$rB))]>;
 }
@@ -3967,7 +3898,6 @@ class FMVecInst<ValueType type>:
 multiclass SFPMul
 {
   def v4f32: FMVecInst<v4f32>;
-  def v2f32: FMVecInst<v2f32>;
   def f32:   FMInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
                      [(set R32FP:$rT, (fmul R32FP:$rA, R32FP:$rB))]>; 
 }
diff --git a/lib/Target/CellSPU/SPUMathInstr.td b/lib/Target/CellSPU/SPUMathInstr.td
index 7205593b04..ed7129e332 100644
--- a/lib/Target/CellSPU/SPUMathInstr.td
+++ b/lib/Target/CellSPU/SPUMathInstr.td
@@ -39,7 +39,7 @@ def : Pat<(mul (v8i16 VECREG:$rA), (v8i16 VECREG:$rB)),
                      (FSMBIv8i16 0xcccc))>;
                  
 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
-// v4i32, v2i32, i32 multiply instruction sequence:
+// v4i32, i32 multiply instruction sequence:
 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
 
 def MPYv4i32:
@@ -49,14 +49,6 @@ def MPYv4i32:
                        (v4i32 (MPYHv4i32 VECREG:$rB, VECREG:$rA)))),
         (v4i32 (MPYUv4i32 VECREG:$rA, VECREG:$rB)))>;
 
-def MPYv2i32:
-  Pat<(mul (v2i32 VECREG:$rA), (v2i32 VECREG:$rB)),
-      (Av2i32
-        (v2i32 (Av2i32 (v2i32 (MPYHv2i32 VECREG:$rA, VECREG:$rB)),
-                       (v2i32 (MPYHv2i32 VECREG:$rB, VECREG:$rA)))),
-        (v2i32 (MPYUv2i32 VECREG:$rA, VECREG:$rB)))>;
-
-
 def MPYi32:
   Pat<(mul R32C:$rA, R32C:$rB),
       (Ar32
diff --git a/lib/Target/CellSPU/SPURegisterInfo.td b/lib/Target/CellSPU/SPURegisterInfo.td
index dd09d50412..3e8f097925 100644
--- a/lib/Target/CellSPU/SPURegisterInfo.td
+++ b/lib/Target/CellSPU/SPURegisterInfo.td
@@ -394,7 +394,7 @@ def R8C : RegisterClass<"SPU", [i8], 128,
 
 // The SPU's registers as vector registers:
 def VECREG : RegisterClass<"SPU",
-                           [v16i8,v8i16,v2i32,v2f32,v4i32,v4f32,v2i64,v2f64],
+                           [v16i8,v8i16,v4i32,v4f32,v2i64,v2f64],
                            128,
  [
    /* volatile register */