diff --git a/llvm/lib/Analysis/ConstantFolding.cpp b/llvm/lib/Analysis/ConstantFolding.cpp
index b504d3310f57..253a7243bcf5 100644
--- a/llvm/lib/Analysis/ConstantFolding.cpp
+++ b/llvm/lib/Analysis/ConstantFolding.cpp
@@ -399,6 +399,12 @@ Constant *llvm::ConstantFoldLoadThroughBitcast(Constant *C, Type *DestTy,
       } while (ElemC && DL.getTypeSizeInBits(ElemC->getType()).isZero());
       C = ElemC;
     } else {
+      // For non-byte-sized vector elements, the first element is not
+      // necessarily located at the vector base address.
+      if (auto *VT = dyn_cast<VectorType>(SrcTy))
+        if (!DL.typeSizeEqualsStoreSize(VT->getElementType()))
+          return nullptr;
+
       C = C->getAggregateElement(0u);
     }
   } while (C);
diff --git a/llvm/test/Transforms/InstCombine/load-store-forward.ll b/llvm/test/Transforms/InstCombine/load-store-forward.ll
index 1f4c49376f9b..c1a01454772f 100644
--- a/llvm/test/Transforms/InstCombine/load-store-forward.ll
+++ b/llvm/test/Transforms/InstCombine/load-store-forward.ll
@@ -47,7 +47,8 @@ define i17 @vec_store_load_first_odd_size(i17* %p) {
 ; CHECK-LABEL: @vec_store_load_first_odd_size(
 ; CHECK-NEXT:    [[P2:%.*]] = bitcast i17* [[P:%.*]] to <2 x i17>*
 ; CHECK-NEXT:    store <2 x i17> <i17 1, i17 2>, <2 x i17>* [[P2]], align 8
-; CHECK-NEXT:    ret i17 1
+; CHECK-NEXT:    [[LOAD:%.*]] = load i17, i17* [[P]], align 4
+; CHECK-NEXT:    ret i17 [[LOAD]]
 ;
   %p2 = bitcast i17* %p to <2 x i17>*
   store <2 x i17> <i17 1, i17 2>, <2 x i17>* %p2