Extend __builtin_shufflevector to expose the full power of the llvm shufflevector instruction. This means it can now be used for vector truncation and concatenation. This will be used for the ARM NEON implementation.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@105589 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 95 insertions, 3 deletions

diff --git a/lib/CodeGen/CGExprScalar.cpp b/lib/CodeGen/CGExprScalar.cpp
index 2108414c5a..0837435316 100644
--- a/lib/CodeGen/CGExprScalar.cpp
+++ b/lib/CodeGen/CGExprScalar.cpp
@@ -578,12 +578,104 @@ Value *ScalarExprEmitter::VisitExpr(Expr *E) {
 }
 
 Value *ScalarExprEmitter::VisitShuffleVectorExpr(ShuffleVectorExpr *E) {
-  llvm::SmallVector<llvm::Constant*, 32> indices;
-  for (unsigned i = 2; i < E->getNumSubExprs(); i++) {
-    indices.push_back(cast<llvm::Constant>(CGF.EmitScalarExpr(E->getExpr(i))));
+  // Vector Mask Case
+  if (E->getNumSubExprs() == 2 || 
+      E->getNumSubExprs() == 3 && E->getExpr(2)->getType()->isVectorType()) {
+    Value* LHS = CGF.EmitScalarExpr(E->getExpr(0));
+    Value* RHS = CGF.EmitScalarExpr(E->getExpr(1));
+    Value* Mask;
+    
+    const llvm::Type *I32Ty = llvm::Type::getInt32Ty(CGF.getLLVMContext());
+    const llvm::VectorType *LTy = cast<llvm::VectorType>(LHS->getType());
+    unsigned LHSElts = LTy->getNumElements();
+
+    if (E->getNumSubExprs() == 3) {
+      Mask = CGF.EmitScalarExpr(E->getExpr(2));
+      
+      // Shuffle LHS & RHS into one input vector.
+      llvm::SmallVector<llvm::Constant*, 32> concat;
+      for (unsigned i = 0; i != LHSElts; ++i) {
+        concat.push_back(llvm::ConstantInt::get(I32Ty, 2*i));
+        concat.push_back(llvm::ConstantInt::get(I32Ty, 2*i+1));
+      }
+      
+      Value* CV = llvm::ConstantVector::get(concat.begin(), concat.size());
+      LHS = Builder.CreateShuffleVector(LHS, RHS, CV, "concat");
+      LHSElts *= 2;
+    } else {
+      Mask = RHS;
+    }
+    
+    const llvm::VectorType *MTy = cast<llvm::VectorType>(Mask->getType());
+    llvm::Constant* EltMask;
+    
+    // Treat vec3 like vec4.
+    if ((LHSElts == 6) && (E->getNumSubExprs() == 3))
+      EltMask = llvm::ConstantInt::get(MTy->getElementType(),
+                                       (1 << llvm::Log2_32(LHSElts+2))-1);
+    else if ((LHSElts == 3) && (E->getNumSubExprs() == 2))
+      EltMask = llvm::ConstantInt::get(MTy->getElementType(),
+                                       (1 << llvm::Log2_32(LHSElts+1))-1);
+    else
+      EltMask = llvm::ConstantInt::get(MTy->getElementType(),
+                                       (1 << llvm::Log2_32(LHSElts))-1);
+             
+    // Mask off the high bits of each shuffle index.
+    llvm::SmallVector<llvm::Constant *, 32> MaskV;
+    for (unsigned i = 0, e = MTy->getNumElements(); i != e; ++i)
+      MaskV.push_back(EltMask);
+    
+    Value* MaskBits = llvm::ConstantVector::get(MaskV.begin(), MaskV.size());
+    Mask = Builder.CreateAnd(Mask, MaskBits, "mask");
+    
+    // newv = undef
+    // mask = mask & maskbits
+    // for each elt
+    //   n = extract mask i
+    //   x = extract val n
+    //   newv = insert newv, x, i
+    const llvm::VectorType *RTy = llvm::VectorType::get(LTy->getElementType(),
+                                                        MTy->getNumElements());
+    Value* NewV = llvm::UndefValue::get(RTy);
+    for (unsigned i = 0, e = MTy->getNumElements(); i != e; ++i) {
+      Value *Indx = llvm::ConstantInt::get(I32Ty, i);
+      Indx = Builder.CreateExtractElement(Mask, Indx, "shuf_idx");
+      Indx = Builder.CreateZExt(Indx, I32Ty, "idx_zext");
+      
+      // Handle vec3 special since the index will be off by one for the RHS.
+      if ((LHSElts == 6) && (E->getNumSubExprs() == 3)) {
+        Value *cmpIndx, *newIndx;
+        cmpIndx = Builder.CreateICmpUGT(Indx, llvm::ConstantInt::get(I32Ty, 3),
+                                        "cmp_shuf_idx");
+        newIndx = Builder.CreateSub(Indx, llvm::ConstantInt::get(I32Ty, 1),
+                                    "shuf_idx_adj");
+        Indx = Builder.CreateSelect(cmpIndx, newIndx, Indx, "sel_shuf_idx");
+      }
+      Value *VExt = Builder.CreateExtractElement(LHS, Indx, "shuf_elt");
+      NewV = Builder.CreateInsertElement(NewV, VExt, Indx, "shuf_ins");
+    }
+    return NewV;
   }
+  
   Value* V1 = CGF.EmitScalarExpr(E->getExpr(0));
   Value* V2 = CGF.EmitScalarExpr(E->getExpr(1));
+  
+  // Handle vec3 special since the index will be off by one for the RHS.
+  llvm::SmallVector<llvm::Constant*, 32> indices;
+  for (unsigned i = 2; i < E->getNumSubExprs(); i++) {
+    llvm::Constant *C = cast<llvm::Constant>(CGF.EmitScalarExpr(E->getExpr(i)));
+    const llvm::VectorType *VTy = cast<llvm::VectorType>(V1->getType());
+    if (VTy->getNumElements() == 3) {
+      if (llvm::ConstantInt *CI = dyn_cast<llvm::ConstantInt>(C)) {
+        uint64_t cVal = CI->getZExtValue();
+        if (cVal > 3) {
+          C = llvm::ConstantInt::get(C->getType(), cVal-1);
+        }
+      }
+    }
+    indices.push_back(C);
+  }
+
   Value* SV = llvm::ConstantVector::get(indices.begin(), indices.size());
   return Builder.CreateShuffleVector(V1, V2, SV, "shuffle");
 }