Index: llvm/include/llvm/Analysis/ConstantFolding.h
===================================================================
--- llvm/include/llvm/Analysis/ConstantFolding.h
+++ llvm/include/llvm/Analysis/ConstantFolding.h
@@ -128,6 +128,17 @@
 Constant *ConstantFoldShuffleVectorInstruction(Constant *V1, Constant *V2,
                                                ArrayRef<int> Mask);
 
+/// Extract value of C at the given Offset reinterpreted as Ty. If bits past
+/// the end of C are accessed, they are assumed to be poison and may take any
+/// value.
+Constant *ConstantFoldLoadFromConst(Constant *C, Type *Ty, const APInt &Offset,
+                                    const DataLayout &DL);
+
+/// Extract value of C reinterpreted as Ty. Same as previous API with zero
+/// offset.
+Constant *ConstantFoldLoadFromConst(Constant *C, Type *Ty,
+                                    const DataLayout &DL);
+
 /// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
 /// produce if it is constant and determinable.  If this is not determinable,
 /// return null.
Index: llvm/lib/Analysis/ConstantFolding.cpp
===================================================================
--- llvm/lib/Analysis/ConstantFolding.cpp
+++ llvm/lib/Analysis/ConstantFolding.cpp
@@ -668,8 +668,11 @@
   return C;
 }
 
-Constant *ConstantFoldLoadFromConst(Constant *C, Type *Ty, const APInt &Offset,
-                                    const DataLayout &DL) {
+} // end anonymous namespace
+
+Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty,
+                                          const APInt &Offset,
+                                          const DataLayout &DL) {
   if (Constant *AtOffset = getConstantAtOffset(C, Offset, DL))
     if (Constant *Result = ConstantFoldLoadThroughBitcast(AtOffset, Ty, DL))
       return Result;
@@ -681,7 +684,10 @@
   return nullptr;
 }
 
-} // end anonymous namespace
+Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty,
+                                          const DataLayout &DL) {
+  return ConstantFoldLoadFromConst(C, Ty, APInt(64, 0), DL);
+}
 
 Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
                                              const DataLayout &DL) {
Index: llvm/lib/Analysis/Loads.cpp
===================================================================
--- llvm/lib/Analysis/Loads.cpp
+++ llvm/lib/Analysis/Loads.cpp
@@ -511,8 +511,11 @@
     if (CastInst::isBitOrNoopPointerCastable(Val->getType(), AccessTy, DL))
       return Val;
 
-    if (auto *C = dyn_cast<Constant>(Val))
-      return ConstantFoldLoadThroughBitcast(C, AccessTy, DL);
+    TypeSize StoreSize = DL.getTypeStoreSize(Val->getType());
+    TypeSize LoadSize = DL.getTypeStoreSize(AccessTy);
+    if (TypeSize::isKnownLE(LoadSize, StoreSize))
+      if (auto *C = dyn_cast<Constant>(Val))
+        return ConstantFoldLoadFromConst(C, AccessTy, DL);
   }
 
   return nullptr;
Index: llvm/test/Transforms/InstCombine/load-store-forward.ll
===================================================================
--- llvm/test/Transforms/InstCombine/load-store-forward.ll
+++ llvm/test/Transforms/InstCombine/load-store-forward.ll
@@ -6,11 +6,13 @@
 ; but is non-trivial.
 
 define i8 @load_smaller_int(i16* %p) {
-; CHECK-LABEL: @load_smaller_int(
-; CHECK-NEXT:    store i16 258, i16* [[P:%.*]], align 2
-; CHECK-NEXT:    [[P2:%.*]] = bitcast i16* [[P]] to i8*
-; CHECK-NEXT:    [[LOAD:%.*]] = load i8, i8* [[P2]], align 1
-; CHECK-NEXT:    ret i8 [[LOAD]]
+; LITTLE-LABEL: @load_smaller_int(
+; LITTLE-NEXT:    store i16 258, i16* [[P:%.*]], align 2
+; LITTLE-NEXT:    ret i8 2
+;
+; BIG-LABEL: @load_smaller_int(
+; BIG-NEXT:    store i16 258, i16* [[P:%.*]], align 2
+; BIG-NEXT:    ret i8 1
 ;
   store i16 258, i16* %p
   %p2 = bitcast i16* %p to i8*