Change handling of illegal vector types to widen when possible instead of

expanding: e.g. <2 x float> -> <4 x float> instead of -> 2 floats.  This
affects two places in the code: handling cross block values and handling
function return and arguments.  Since vectors are already widened by 
legalizetypes, this gives us much better code and unblocks x86-64 abi
and SPU abi work.

For example, this (which is a silly example of a cross-block value):
define <4 x float> @test2(<4 x float> %A) nounwind {
 %B = shufflevector <4 x float> %A, <4 x float> undef, <2 x i32> <i32 0, i32 1>
 %C = fadd <2 x float> %B, %B
  br label %BB
BB:
 %D = fadd <2 x float> %C, %C
 %E = shufflevector <2 x float> %D, <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
 ret <4 x float> %E
}

Now compiles into:

_test2:                                 ## @test2
## BB#0:
 addps %xmm0, %xmm0
 addps %xmm0, %xmm0
 ret

previously it compiled into:

_test2:                                 ## @test2
## BB#0:
 addps %xmm0, %xmm0
 pshufd $1, %xmm0, %xmm1
                                        ## kill: XMM0<def> XMM0<kill> XMM0<def>
 insertps $0, %xmm0, %xmm0
 insertps $16, %xmm1, %xmm0
 addps %xmm0, %xmm0
 ret

This implements rdar://8230384



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@112101 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 97 insertions, 46 deletions

diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index a8a8ecb..35e7520 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -252,8 +252,21 @@ static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
   if (PartVT == ValueVT)
     return Val;
   
-  if (PartVT.isVector())  // Vector/Vector bitcast.
+  if (PartVT.isVector()) {
+    // If the element type of the source/dest vectors are the same, but the
+    // parts vector has more elements than the value vector, then we have a
+    // vector widening case (e.g. <2 x float> -> <4 x float>).  Extract the
+    // elements we want.
+    if (PartVT.getVectorElementType() == ValueVT.getVectorElementType()) {
+      assert(PartVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
+             "Cannot narrow, it would be a lossy transformation");
+      return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
+                         DAG.getIntPtrConstant(0));
+    }                                      
+    
+    // Vector/Vector bitcast.
     return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val);
+  }
   
   assert(ValueVT.getVectorElementType() == PartVT &&
          ValueVT.getVectorNumElements() == 1 &&
@@ -392,16 +405,39 @@ static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL,
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   
   if (NumParts == 1) {
-    if (PartVT != ValueVT) {
-      if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
-        Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val);
-      } else {
-        assert(ValueVT.getVectorElementType() == PartVT &&
-               ValueVT.getVectorNumElements() == 1 &&
-               "Only trivial vector-to-scalar conversions should get here!");
-        Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
-                          PartVT, Val, DAG.getIntPtrConstant(0));
-      }
+    if (PartVT == ValueVT) {
+      // Nothing to do.
+    } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
+      // Bitconvert vector->vector case.
+      Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val);
+    } else if (PartVT.isVector() &&
+               PartVT.getVectorElementType() == ValueVT.getVectorElementType()&&
+               PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
+      EVT ElementVT = PartVT.getVectorElementType();
+      // Vector widening case, e.g. <2 x float> -> <4 x float>.  Shuffle in
+      // undef elements.
+      SmallVector<SDValue, 16> Ops;
+      for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
+        Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
+                                  ElementVT, Val, DAG.getIntPtrConstant(i)));
+      
+      for (unsigned i = ValueVT.getVectorNumElements(),
+           e = PartVT.getVectorNumElements(); i != e; ++i)
+        Ops.push_back(DAG.getUNDEF(ElementVT));
+
+      Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size());
+
+      // FIXME: Use CONCAT for 2x -> 4x.
+      
+      //SDValue UndefElts = DAG.getUNDEF(VectorTy);
+      //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
+    } else {
+      // Vector -> scalar conversion.
+      assert(ValueVT.getVectorElementType() == PartVT &&
+             ValueVT.getVectorNumElements() == 1 &&
+             "Only trivial vector-to-scalar conversions should get here!");
+      Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
+                        PartVT, Val, DAG.getIntPtrConstant(0));
     }
     
     Parts[0] = Val;
@@ -428,8 +464,7 @@ static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL,
                    DAG.getIntPtrConstant(i * (NumElements / NumIntermediates)));
     else
       Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
-                           IntermediateVT, Val,
-                           DAG.getIntPtrConstant(i));
+                           IntermediateVT, Val, DAG.getIntPtrConstant(i));
   }
   
   // Split the intermediate operands into legal parts.
diff --git a/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 9772d55..077fd1d 100644
--- a/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -697,6 +697,7 @@ TargetLowering::findRepresentativeClass(EVT VT) const {
   return std::make_pair(BestRC, 1);
 }
 
+
 /// computeRegisterProperties - Once all of the register classes are added,
 /// this allows us to compute derived properties we expose.
 void TargetLowering::computeRegisterProperties() {
@@ -782,6 +783,28 @@ void TargetLowering::computeRegisterProperties() {
     MVT VT = (MVT::SimpleValueType)i;
     if (isTypeLegal(VT)) continue;
     
+    // Determine if there is a legal wider type.  If so, we should promote to
+    // that wider vector type.
+    EVT EltVT = VT.getVectorElementType();
+    unsigned NElts = VT.getVectorNumElements();
+    if (NElts != 1) {
+      bool IsLegalWiderType = false;
+      for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
+        EVT SVT = (MVT::SimpleValueType)nVT;
+        if (SVT.getVectorElementType() == EltVT &&
+            SVT.getVectorNumElements() > NElts && 
+            isTypeSynthesizable(SVT)) {
+          TransformToType[i] = SVT;
+          RegisterTypeForVT[i] = SVT;
+          NumRegistersForVT[i] = 1;
+          ValueTypeActions.setTypeAction(VT, Promote);
+          IsLegalWiderType = true;
+          break;
+        }
+      }
+      if (IsLegalWiderType) continue;
+    }
+    
     MVT IntermediateVT;
     EVT RegisterVT;
     unsigned NumIntermediates;
@@ -790,30 +813,14 @@ void TargetLowering::computeRegisterProperties() {
                                 RegisterVT, this);
     RegisterTypeForVT[i] = RegisterVT;
     
-    // Determine if there is a legal wider type.
-    bool IsLegalWiderType = false;
-    EVT EltVT = VT.getVectorElementType();
-    unsigned NElts = VT.getVectorNumElements();
-    for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
-      EVT SVT = (MVT::SimpleValueType)nVT;
-      if (isTypeSynthesizable(SVT) && SVT.getVectorElementType() == EltVT &&
-          SVT.getVectorNumElements() > NElts && NElts != 1) {
-        TransformToType[i] = SVT;
-        ValueTypeActions.setTypeAction(VT, Promote);
-        IsLegalWiderType = true;
-        break;
-      }
-    }
-    if (!IsLegalWiderType) {
-      EVT NVT = VT.getPow2VectorType();
-      if (NVT == VT) {
-        // Type is already a power of 2.  The default action is to split.
-        TransformToType[i] = MVT::Other;
-        ValueTypeActions.setTypeAction(VT, Expand);
-      } else {
-        TransformToType[i] = NVT;
-        ValueTypeActions.setTypeAction(VT, Promote);
-      }
+    EVT NVT = VT.getPow2VectorType();
+    if (NVT == VT) {
+      // Type is already a power of 2.  The default action is to split.
+      TransformToType[i] = MVT::Other;
+      ValueTypeActions.setTypeAction(VT, Expand);
+    } else {
+      TransformToType[i] = NVT;
+      ValueTypeActions.setTypeAction(VT, Promote);
     }
   }
 
@@ -857,8 +864,21 @@ unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
                                                 EVT &IntermediateVT,
                                                 unsigned &NumIntermediates,
                                                 EVT &RegisterVT) const {
-  // Figure out the right, legal destination reg to copy into.
   unsigned NumElts = VT.getVectorNumElements();
+  
+  // If there is a wider vector type with the same element type as this one,
+  // we should widen to that legal vector type.  This handles things like
+  // <2 x float> -> <4 x float>.
+  if (NumElts != 1 && getTypeAction(Context, VT) == Promote) {
+    RegisterVT = getTypeToTransformTo(Context, VT);
+    if (isTypeLegal(RegisterVT)) {
+      IntermediateVT = RegisterVT;
+      NumIntermediates = 1;
+      return 1;
+    }
+  }
+  
+  // Figure out the right, legal destination reg to copy into.
   EVT EltTy = VT.getVectorElementType();
   
   unsigned NumVectorRegs = 1;
@@ -887,16 +907,12 @@ unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
 
   EVT DestVT = getRegisterType(Context, NewVT);
   RegisterVT = DestVT;
-  if (DestVT.bitsLT(NewVT)) {
-    // Value is expanded, e.g. i64 -> i16.
+  if (DestVT.bitsLT(NewVT))   // Value is expanded, e.g. i64 -> i16.
     return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits());
-  } else {
-    // Otherwise, promotion or legal types use the same number of registers as
-    // the vector decimated to the appropriate level.
-    return NumVectorRegs;
-  }
   
-  return 1;
+  // Otherwise, promotion or legal types use the same number of registers as
+  // the vector decimated to the appropriate level.
+  return NumVectorRegs;
 }
 
 /// Get the EVTs and ArgFlags collections that represent the legalized return