Index: lib/Transforms/InstCombine/InstCombineCalls.cpp
===================================================================
--- lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -136,6 +136,14 @@
   if (Size > 8 || (Size&(Size-1)))
     return nullptr;  // If not 1/2/4/8 bytes, exit.
 
+  // If it is an atomic and alignment is less than the size then we will
+  // introduce the unaligned memory access which will be later transformed
+  // into libcall in CodeGen. This is not evident performance gain so disable
+  // it now.
+  if (isa<AtomicMemTransferInst>(MI))
+    if (CopyDstAlign < Size || CopySrcAlign < Size)
+      return nullptr;
+
   // Use an integer load+store unless we can find something better.
   unsigned SrcAddrSp =
     cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace();
@@ -220,6 +228,18 @@
   Alignment = MI->getDestAlignment();
   assert(Len && "0-sized memory setting should be removed already.");
 
+  // Alignment 0 is identity for alignment 1 for memset, but not store.
+  if (Alignment == 0)
+    Alignment = 1;
+
+  // If it is an atomic and alignment is less than the size then we will
+  // introduce the unaligned memory access which will be later transformed
+  // into libcall in CodeGen. This is not evident performance gain so disable
+  // it now.
+  if (isa<AtomicMemSetInst>(MI))
+    if (Alignment < Len)
+      return nullptr;
+
   // memset(s,c,n) -> store s, c (for n=1,2,4,8)
   if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
     Type *ITy = IntegerType::get(MI->getContext(), Len*8);  // n=1 -> i8.
@@ -229,9 +249,6 @@
     Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp);
     Dest = Builder.CreateBitCast(Dest, NewDstPtrTy);
 
-    // Alignment 0 is identity for alignment 1 for memset, but not store.
-    if (Alignment == 0) Alignment = 1;
-
     // Extract the fill value and store.
     uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
     StoreInst *S = Builder.CreateStore(ConstantInt::get(ITy, Fill), Dest,
Index: test/Transforms/InstCombine/element-atomic-memintrins.ll
===================================================================
--- test/Transforms/InstCombine/element-atomic-memintrins.ll
+++ test/Transforms/InstCombine/element-atomic-memintrins.ll
@@ -15,12 +15,9 @@
 define void @test_memset_to_store(i8* %dest) {
 ; CHECK-LABEL: @test_memset_to_store(
 ; CHECK-NEXT:    store atomic i8 1, i8* [[DEST:%.*]] unordered, align 1
-; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i8* [[DEST]] to i16*
-; CHECK-NEXT:    store atomic i16 257, i16* [[TMP1]] unordered, align 1
-; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[DEST]] to i32*
-; CHECK-NEXT:    store atomic i32 16843009, i32* [[TMP2]] unordered, align 1
-; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i8* [[DEST]] to i64*
-; CHECK-NEXT:    store atomic i64 72340172838076673, i64* [[TMP3]] unordered, align 1
+; CHECK-NEXT:    call void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* align 1 [[DEST]], i8 1, i32 2, i32 1)
+; CHECK-NEXT:    call void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* align 1 [[DEST]], i8 1, i32 4, i32 1)
+; CHECK-NEXT:    call void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* align 1 [[DEST]], i8 1, i32 8, i32 1)
 ; CHECK-NEXT:    call void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* align 1 [[DEST]], i8 1, i32 16, i32 1)
 ; CHECK-NEXT:    ret void
 ;
@@ -36,10 +33,8 @@
 ; CHECK-LABEL: @test_memset_to_store_2(
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i8* [[DEST:%.*]] to i16*
 ; CHECK-NEXT:    store atomic i16 257, i16* [[TMP1]] unordered, align 2
-; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[DEST]] to i32*
-; CHECK-NEXT:    store atomic i32 16843009, i32* [[TMP2]] unordered, align 2
-; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i8* [[DEST]] to i64*
-; CHECK-NEXT:    store atomic i64 72340172838076673, i64* [[TMP3]] unordered, align 2
+; CHECK-NEXT:    call void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* align 2 [[DEST]], i8 1, i32 4, i32 2)
+; CHECK-NEXT:    call void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* align 2 [[DEST]], i8 1, i32 8, i32 2)
 ; CHECK-NEXT:    call void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* align 2 [[DEST]], i8 1, i32 16, i32 2)
 ; CHECK-NEXT:    ret void
 ;
@@ -54,8 +49,7 @@
 ; CHECK-LABEL: @test_memset_to_store_4(
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i8* [[DEST:%.*]] to i32*
 ; CHECK-NEXT:    store atomic i32 16843009, i32* [[TMP1]] unordered, align 4
-; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[DEST]] to i64*
-; CHECK-NEXT:    store atomic i64 72340172838076673, i64* [[TMP2]] unordered, align 4
+; CHECK-NEXT:    call void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* align 4 [[DEST]], i8 1, i32 8, i32 4)
 ; CHECK-NEXT:    call void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* align 4 [[DEST]], i8 1, i32 16, i32 4)
 ; CHECK-NEXT:    ret void
 ;
@@ -134,18 +128,9 @@
 ; CHECK-LABEL: @test_memmove_loadstore(
 ; CHECK-NEXT:    [[TMP1:%.*]] = load atomic i8, i8* [[SRC:%.*]] unordered, align 1
 ; CHECK-NEXT:    store atomic i8 [[TMP1]], i8* [[DEST:%.*]] unordered, align 1
-; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[SRC]] to i16*
-; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i8* [[DEST]] to i16*
-; CHECK-NEXT:    [[TMP4:%.*]] = load atomic i16, i16* [[TMP2]] unordered, align 1
-; CHECK-NEXT:    store atomic i16 [[TMP4]], i16* [[TMP3]] unordered, align 1
-; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i8* [[SRC]] to i32*
-; CHECK-NEXT:    [[TMP6:%.*]] = bitcast i8* [[DEST]] to i32*
-; CHECK-NEXT:    [[TMP7:%.*]] = load atomic i32, i32* [[TMP5]] unordered, align 1
-; CHECK-NEXT:    store atomic i32 [[TMP7]], i32* [[TMP6]] unordered, align 1
-; CHECK-NEXT:    [[TMP8:%.*]] = bitcast i8* [[SRC]] to i64*
-; CHECK-NEXT:    [[TMP9:%.*]] = bitcast i8* [[DEST]] to i64*
-; CHECK-NEXT:    [[TMP10:%.*]] = load atomic i64, i64* [[TMP8]] unordered, align 1
-; CHECK-NEXT:    store atomic i64 [[TMP10]], i64* [[TMP9]] unordered, align 1
+; CHECK-NEXT:    call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* align 1 [[DEST]], i8* align 1 [[SRC]], i32 2, i32 1)
+; CHECK-NEXT:    call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* align 1 [[DEST]], i8* align 1 [[SRC]], i32 4, i32 1)
+; CHECK-NEXT:    call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* align 1 [[DEST]], i8* align 1 [[SRC]], i32 8, i32 1)
 ; CHECK-NEXT:    call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* align 1 [[DEST]], i8* align 1 [[SRC]], i32 16, i32 1)
 ; CHECK-NEXT:    ret void
 ;
@@ -163,14 +148,8 @@
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[DEST:%.*]] to i16*
 ; CHECK-NEXT:    [[TMP3:%.*]] = load atomic i16, i16* [[TMP1]] unordered, align 2
 ; CHECK-NEXT:    store atomic i16 [[TMP3]], i16* [[TMP2]] unordered, align 2
-; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i8* [[SRC]] to i32*
-; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i8* [[DEST]] to i32*
-; CHECK-NEXT:    [[TMP6:%.*]] = load atomic i32, i32* [[TMP4]] unordered, align 2
-; CHECK-NEXT:    store atomic i32 [[TMP6]], i32* [[TMP5]] unordered, align 2
-; CHECK-NEXT:    [[TMP7:%.*]] = bitcast i8* [[SRC]] to i64*
-; CHECK-NEXT:    [[TMP8:%.*]] = bitcast i8* [[DEST]] to i64*
-; CHECK-NEXT:    [[TMP9:%.*]] = load atomic i64, i64* [[TMP7]] unordered, align 2
-; CHECK-NEXT:    store atomic i64 [[TMP9]], i64* [[TMP8]] unordered, align 2
+; CHECK-NEXT:    call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* align 2 [[DEST]], i8* align 2 [[SRC]], i32 4, i32 2)
+; CHECK-NEXT:    call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* align 2 [[DEST]], i8* align 2 [[SRC]], i32 8, i32 2)
 ; CHECK-NEXT:    call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* align 2 [[DEST]], i8* align 2 [[SRC]], i32 16, i32 2)
 ; CHECK-NEXT:    ret void
 ;
@@ -187,10 +166,7 @@
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[DEST:%.*]] to i32*
 ; CHECK-NEXT:    [[TMP3:%.*]] = load atomic i32, i32* [[TMP1]] unordered, align 4
 ; CHECK-NEXT:    store atomic i32 [[TMP3]], i32* [[TMP2]] unordered, align 4
-; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i8* [[SRC]] to i64*
-; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i8* [[DEST]] to i64*
-; CHECK-NEXT:    [[TMP6:%.*]] = load atomic i64, i64* [[TMP4]] unordered, align 4
-; CHECK-NEXT:    store atomic i64 [[TMP6]], i64* [[TMP5]] unordered, align 4
+; CHECK-NEXT:    call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 [[DEST]], i8* align 4 [[SRC]], i32 8, i32 4)
 ; CHECK-NEXT:    call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 [[DEST]], i8* align 4 [[SRC]], i32 16, i32 4)
 ; CHECK-NEXT:    ret void
 ;
@@ -258,18 +234,9 @@
 ; CHECK-LABEL: @test_memcpy_loadstore(
 ; CHECK-NEXT:    [[TMP1:%.*]] = load atomic i8, i8* [[SRC:%.*]] unordered, align 1
 ; CHECK-NEXT:    store atomic i8 [[TMP1]], i8* [[DEST:%.*]] unordered, align 1
-; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[SRC]] to i16*
-; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i8* [[DEST]] to i16*
-; CHECK-NEXT:    [[TMP4:%.*]] = load atomic i16, i16* [[TMP2]] unordered, align 1
-; CHECK-NEXT:    store atomic i16 [[TMP4]], i16* [[TMP3]] unordered, align 1
-; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i8* [[SRC]] to i32*
-; CHECK-NEXT:    [[TMP6:%.*]] = bitcast i8* [[DEST]] to i32*
-; CHECK-NEXT:    [[TMP7:%.*]] = load atomic i32, i32* [[TMP5]] unordered, align 1
-; CHECK-NEXT:    store atomic i32 [[TMP7]], i32* [[TMP6]] unordered, align 1
-; CHECK-NEXT:    [[TMP8:%.*]] = bitcast i8* [[SRC]] to i64*
-; CHECK-NEXT:    [[TMP9:%.*]] = bitcast i8* [[DEST]] to i64*
-; CHECK-NEXT:    [[TMP10:%.*]] = load atomic i64, i64* [[TMP8]] unordered, align 1
-; CHECK-NEXT:    store atomic i64 [[TMP10]], i64* [[TMP9]] unordered, align 1
+; CHECK-NEXT:    call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 1 [[DEST]], i8* align 1 [[SRC]], i32 2, i32 1)
+; CHECK-NEXT:    call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 1 [[DEST]], i8* align 1 [[SRC]], i32 4, i32 1)
+; CHECK-NEXT:    call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 1 [[DEST]], i8* align 1 [[SRC]], i32 8, i32 1)
 ; CHECK-NEXT:    call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 1 [[DEST]], i8* align 1 [[SRC]], i32 16, i32 1)
 ; CHECK-NEXT:    ret void
 ;
@@ -287,14 +254,8 @@
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[DEST:%.*]] to i16*
 ; CHECK-NEXT:    [[TMP3:%.*]] = load atomic i16, i16* [[TMP1]] unordered, align 2
 ; CHECK-NEXT:    store atomic i16 [[TMP3]], i16* [[TMP2]] unordered, align 2
-; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i8* [[SRC]] to i32*
-; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i8* [[DEST]] to i32*
-; CHECK-NEXT:    [[TMP6:%.*]] = load atomic i32, i32* [[TMP4]] unordered, align 2
-; CHECK-NEXT:    store atomic i32 [[TMP6]], i32* [[TMP5]] unordered, align 2
-; CHECK-NEXT:    [[TMP7:%.*]] = bitcast i8* [[SRC]] to i64*
-; CHECK-NEXT:    [[TMP8:%.*]] = bitcast i8* [[DEST]] to i64*
-; CHECK-NEXT:    [[TMP9:%.*]] = load atomic i64, i64* [[TMP7]] unordered, align 2
-; CHECK-NEXT:    store atomic i64 [[TMP9]], i64* [[TMP8]] unordered, align 2
+; CHECK-NEXT:    call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 2 [[DEST]], i8* align 2 [[SRC]], i32 4, i32 2)
+; CHECK-NEXT:    call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 2 [[DEST]], i8* align 2 [[SRC]], i32 8, i32 2)
 ; CHECK-NEXT:    call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 2 [[DEST]], i8* align 2 [[SRC]], i32 16, i32 2)
 ; CHECK-NEXT:    ret void
 ;
@@ -311,10 +272,7 @@
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[DEST:%.*]] to i32*
 ; CHECK-NEXT:    [[TMP3:%.*]] = load atomic i32, i32* [[TMP1]] unordered, align 4
 ; CHECK-NEXT:    store atomic i32 [[TMP3]], i32* [[TMP2]] unordered, align 4
-; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i8* [[SRC]] to i64*
-; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i8* [[DEST]] to i64*
-; CHECK-NEXT:    [[TMP6:%.*]] = load atomic i64, i64* [[TMP4]] unordered, align 4
-; CHECK-NEXT:    store atomic i64 [[TMP6]], i64* [[TMP5]] unordered, align 4
+; CHECK-NEXT:    call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 [[DEST]], i8* align 4 [[SRC]], i32 8, i32 4)
 ; CHECK-NEXT:    call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 [[DEST]], i8* align 4 [[SRC]], i32 16, i32 4)
 ; CHECK-NEXT:    ret void
 ;