diff --git a/lib/DXIL/DxilResource.cpp b/lib/DXIL/DxilResource.cpp
index 7124417dcf..3ab71030bb 100644
--- a/lib/DXIL/DxilResource.cpp
+++ b/lib/DXIL/DxilResource.cpp
@@ -30,18 +30,7 @@ DxilResource::DxilResource()
 
 CompType DxilResource::GetCompType() const { return m_CompType; }
 
-void DxilResource::SetCompType(const CompType CT) {
-  // Translate packed types to u32
-  switch (CT.GetKind()) {
-  case CompType::Kind::PackedS8x32:
-  case CompType::Kind::PackedU8x32:
-    m_CompType = CompType::getU32();
-    break;
-  default:
-    m_CompType = CT;
-    break;
-  }
-}
+void DxilResource::SetCompType(const CompType CT) { m_CompType = CT; }
 
 Type *DxilResource::GetRetType() const {
   Type *Ty = GetHLSLType()->getPointerElementType();
diff --git a/lib/HLSL/HLOperationLower.cpp b/lib/HLSL/HLOperationLower.cpp
index f7b0241ad8..82d59ed806 100644
--- a/lib/HLSL/HLOperationLower.cpp
+++ b/lib/HLSL/HLOperationLower.cpp
@@ -8436,89 +8436,52 @@ void TranslateStructBufSubscriptUser(Instruction *user, Value *handle,
                                    : stInst->getValueOperand()->getType();
     Type *pOverloadTy = Ty->getScalarType();
     Value *offset = baseOffset;
-    unsigned arraySize = 1;
-    Value *eltSize = nullptr;
 
-    if (pOverloadTy->isArrayTy()) {
-      arraySize = pOverloadTy->getArrayNumElements();
-      eltSize = OP->GetU32Const(
-          DL.getTypeAllocSize(pOverloadTy->getArrayElementType()));
+    if (ldInst) {
+      unsigned numComponents = 0;
+      Value *newLd = nullptr;
+      if (VectorType *VTy = dyn_cast<VectorType>(Ty))
+        numComponents = VTy->getNumElements();
+      else
+        numComponents = 1;
 
-      pOverloadTy = pOverloadTy->getArrayElementType()->getScalarType();
-    }
+      if (ResKind == HLResource::Kind::TypedBuffer) {
+        // Typed buffer cannot have offsets, they must be loaded all at once
+        ResRetValueArray ResRet = GenerateTypedBufferLoad(
+            handle, pOverloadTy, bufIdx, status, OP, Builder);
 
-    if (ldInst) {
-      auto LdElement = [=](Value *offset, IRBuilder<> &Builder) -> Value * {
-        unsigned numComponents = 0;
-        if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
-          numComponents = VTy->getNumElements();
-        } else {
-          numComponents = 1;
-        }
+        newLd = ExtractFromTypedBufferLoad(ResRet, Ty, offset, Builder);
+      } else {
+        Value *ResultElts[4];
         Constant *alignment =
             OP->GetI32Const(DL.getTypeAllocSize(Ty->getScalarType()));
-        if (ResKind == HLResource::Kind::TypedBuffer) {
-          // Typed buffer cannot have offsets, they must be loaded all at once
-          ResRetValueArray ResRet = GenerateTypedBufferLoad(
-              handle, pOverloadTy, bufIdx, status, OP, Builder);
-
-          return ExtractFromTypedBufferLoad(ResRet, Ty, offset, Builder);
-        } else {
-          Value *ResultElts[4];
-          GenerateRawBufLd(handle, bufIdx, offset, status, pOverloadTy,
-                           ResultElts, OP, Builder, numComponents, alignment);
-          return ScalarizeElements(Ty, ResultElts, Builder);
-        }
-      };
-
-      Value *newLd = LdElement(offset, Builder);
-      if (arraySize > 1) {
-        newLd =
-            Builder.CreateInsertValue(UndefValue::get(Ty), newLd, (uint64_t)0);
-
-        for (unsigned i = 1; i < arraySize; i++) {
-          offset = Builder.CreateAdd(offset, eltSize);
-          Value *eltLd = LdElement(offset, Builder);
-          newLd = Builder.CreateInsertValue(newLd, eltLd, i);
-        }
+        GenerateRawBufLd(handle, bufIdx, offset, status, pOverloadTy,
+                         ResultElts, OP, Builder, numComponents, alignment);
+        newLd = ScalarizeElements(Ty, ResultElts, Builder);
       }
+
       ldInst->replaceAllUsesWith(newLd);
     } else {
       Value *val = stInst->getValueOperand();
-      auto StElement = [&](Value *offset, Value *val, IRBuilder<> &Builder) {
-        Value *undefVal = llvm::UndefValue::get(pOverloadTy);
-        Value *vals[] = {undefVal, undefVal, undefVal, undefVal};
-        uint8_t mask = 0;
-        if (Ty->isVectorTy()) {
-          unsigned vectorNumElements = Ty->getVectorNumElements();
-          DXASSERT(vectorNumElements <= 4, "up to 4 elements in vector");
-          assert(vectorNumElements <= 4);
-          for (unsigned i = 0; i < vectorNumElements; i++) {
-            vals[i] = Builder.CreateExtractElement(val, i);
-            mask |= (1 << i);
-          }
-        } else {
-          vals[0] = val;
-          mask = DXIL::kCompMask_X;
-        }
-        Constant *alignment =
-            OP->GetI32Const(DL.getTypeAllocSize(Ty->getScalarType()));
-        GenerateStructBufSt(handle, bufIdx, offset, pOverloadTy, OP, Builder,
-                            vals, mask, alignment);
-      };
-      if (arraySize > 1)
-        val = Builder.CreateExtractValue(val, 0);
-
-      StElement(offset, val, Builder);
-      if (arraySize > 1) {
-        val = stInst->getValueOperand();
-
-        for (unsigned i = 1; i < arraySize; i++) {
-          offset = Builder.CreateAdd(offset, eltSize);
-          Value *eltVal = Builder.CreateExtractValue(val, i);
-          StElement(offset, eltVal, Builder);
+      Value *undefVal = llvm::UndefValue::get(pOverloadTy);
+      Value *vals[] = {undefVal, undefVal, undefVal, undefVal};
+      uint8_t mask = 0;
+      if (Ty->isVectorTy()) {
+        unsigned vectorNumElements = Ty->getVectorNumElements();
+        DXASSERT(vectorNumElements <= 4, "up to 4 elements in vector");
+        assert(vectorNumElements <= 4);
+        for (unsigned i = 0; i < vectorNumElements; i++) {
+          vals[i] = Builder.CreateExtractElement(val, i);
+          mask |= (1 << i);
         }
+      } else {
+        vals[0] = val;
+        mask = DXIL::kCompMask_X;
       }
+      Constant *alignment =
+          OP->GetI32Const(DL.getTypeAllocSize(Ty->getScalarType()));
+      GenerateStructBufSt(handle, bufIdx, offset, pOverloadTy, OP, Builder,
+                          vals, mask, alignment);
     }
     user->eraseFromParent();
   } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(user)) {
diff --git a/tools/clang/lib/CodeGen/CGHLSLMS.cpp b/tools/clang/lib/CodeGen/CGHLSLMS.cpp
index 99845893b2..373da18ecd 100644
--- a/tools/clang/lib/CodeGen/CGHLSLMS.cpp
+++ b/tools/clang/lib/CodeGen/CGHLSLMS.cpp
@@ -3539,11 +3539,20 @@ bool CGMSHLSLRuntime::SetUAVSRV(SourceLocation loc,
     if (const BuiltinType *BTy = EltTy->getAs<BuiltinType>()) {
       CompType::Kind kind = BuiltinTyToCompTy(BTy, bHasNormAttribute && bSNorm,
                                               bHasNormAttribute && !bSNorm);
-      // 64bits types are implemented with u32.
-      if (kind == CompType::Kind::U64 || kind == CompType::Kind::I64 ||
-          kind == CompType::Kind::SNormF64 ||
-          kind == CompType::Kind::UNormF64 || kind == CompType::Kind::F64) {
+      // Boolean, 64-bit, and packed types are implemented with u32.
+      switch (kind) {
+      case CompType::Kind::I1:
+      case CompType::Kind::U64:
+      case CompType::Kind::I64:
+      case CompType::Kind::F64:
+      case CompType::Kind::SNormF64:
+      case CompType::Kind::UNormF64:
+      case CompType::Kind::PackedS8x32:
+      case CompType::Kind::PackedU8x32:
         kind = CompType::Kind::U32;
+        break;
+      default:
+        break;
       }
       hlslRes->SetCompType(kind);
     } else {
diff --git a/tools/clang/test/CodeGenDXIL/hlsl/intrinsics/buffer-load-stores.hlsl b/tools/clang/test/CodeGenDXIL/hlsl/intrinsics/buffer-load-stores.hlsl
index 70f9e106e3..82781c2c04 100644
--- a/tools/clang/test/CodeGenDXIL/hlsl/intrinsics/buffer-load-stores.hlsl
+++ b/tools/clang/test/CodeGenDXIL/hlsl/intrinsics/buffer-load-stores.hlsl
@@ -1,10 +1,10 @@
 // RUN: %dxc -DTYPE=float4    -DIX=SIx -T vs_6_6 %s | FileCheck %s
-// RUiN: %dxc -DTYPE=bool4     -DIX=SIx -T vs_6_6 %s | FileCheck %s --check-prefixes=CHECK,I1
+// RUN: %dxc -DTYPE=bool4     -DIX=SIx -T vs_6_6 %s | FileCheck %s --check-prefixes=CHECK,I1
 // RUN: %dxc -DTYPE=uint64_t2 -DIX=SIx -T vs_6_6 %s | FileCheck %s --check-prefixes=CHECK,I64
 // RUN: %dxc -DTYPE=double2   -DIX=SIx -T vs_6_6 %s | FileCheck %s --check-prefixes=CHECK,F64
 
 // RUN: %dxc -DTYPE=float4    -DIX=VIx -T vs_6_6 %s | FileCheck %s --check-prefixes=CHECK,VIX
-// RUiN: %dxc -DTYPE=bool4     -DIX=VIx -T vs_6_6 %s | FileCheck %s --check-prefixes=CHECK,VIX,I1
+// RUN: %dxc -DTYPE=bool4     -DIX=VIx -T vs_6_6 %s | FileCheck %s --check-prefixes=CHECK,VIX,I1
 // RUN: %dxc -DTYPE=uint64_t2 -DIX=VIx -T vs_6_6 %s | FileCheck %s --check-prefixes=CHECK,VIX,I64
 // RUN: %dxc -DTYPE=double2   -DIX=VIx -T vs_6_6 %s | FileCheck %s --check-prefixes=CHECK,VIX,F64
 
diff --git a/tools/clang/test/CodeGenHLSL/dot.hlsl b/tools/clang/test/CodeGenDXIL/hlsl/intrinsics/dot.hlsl
similarity index 100%
rename from tools/clang/test/CodeGenHLSL/dot.hlsl
rename to tools/clang/test/CodeGenDXIL/hlsl/intrinsics/dot.hlsl
diff --git a/tools/clang/test/CodeGenHLSL/vector-and.hlsl b/tools/clang/test/CodeGenDXIL/hlsl/intrinsics/vector-and.hlsl
similarity index 100%
rename from tools/clang/test/CodeGenHLSL/vector-and.hlsl
rename to tools/clang/test/CodeGenDXIL/hlsl/intrinsics/vector-and.hlsl
diff --git a/tools/clang/test/CodeGenHLSL/vector-or.hlsl b/tools/clang/test/CodeGenDXIL/hlsl/intrinsics/vector-or.hlsl
similarity index 100%
rename from tools/clang/test/CodeGenHLSL/vector-or.hlsl
rename to tools/clang/test/CodeGenDXIL/hlsl/intrinsics/vector-or.hlsl