Index: lib/Transforms/InstCombine/InstCombineCalls.cpp
===================================================================
--- lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -133,6 +133,16 @@
   uint64_t Size = MemOpLength->getLimitedValue();
   assert(Size && "0-sized memory transferring should be removed already.");
 
+  // Do not expand a 8 byte copy to load/stores when we optimise for minimum
+  // code size. This could for example expand into 2 word loads and
+  // 2 stores. But when unaligned data access is not supported, this is a
+  // lot worse and we will have 8 byte loads and 8 byte stores. Keeping the
+  // memcpy call will result in just 2 instruction: the call and a mov imm to
+  // an arg register for the number of bytes to copy.
+  auto F = MI->getParent()->getParent();
+  if (Size > 4 && F->hasFnAttribute(Attribute::MinSize))
+    return nullptr;
+
   if (Size > 8 || (Size&(Size-1)))
     return nullptr;  // If not 1/2/4/8 bytes, exit.
 
Index: test/Transforms/InstCombine/memcpy-to-load.ll
===================================================================
--- test/Transforms/InstCombine/memcpy-to-load.ll
+++ test/Transforms/InstCombine/memcpy-to-load.ll
@@ -76,6 +76,16 @@
   ret void
 }
 
+; Do not expand the call if we optimise for minsize
+define void @copy_8_bytes_minsize(i8* %A, i8* %B) #0 {
+; CHECK-LABEL: @test4(
+; CHECK:         tail call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 1 %A, i8* align 1 %B, i32 8, i1 false)
+; CHECK-NEXT:    ret void
+entry:
+  tail call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 1 %A, i8* align 1 %B, i32 8, i1 false)
+  ret void
+}
+
 define void @copy_16_bytes(i8* %d, i8* %s) {
 ; ALL-LABEL: @copy_16_bytes(
 ; ALL-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 1 [[D:%.*]], i8* align 1 [[S:%.*]], i32 16, i1 false)
@@ -85,3 +95,5 @@
   ret void
 }
 
+attributes #0 = { minsize nounwind optsize }
+