diff --git a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
--- a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
+++ b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
@@ -1152,8 +1152,14 @@
   // still want to eliminate the intermediate value, but we have to generate a
   // memmove instead of memcpy.
   bool UseMemMove = false;
-  if (isModSet(BAA.getModRefInfo(M, MemoryLocation::getForSource(MDep))))
+  if (isModSet(BAA.getModRefInfo(M, MemoryLocation::getForSource(MDep)))) {
+    // Don't convert llvm.memcpy.inline into memmove because memmove can be
+    // lowered as a call, and that is not allowed for llvm.memcpy.inline (and
+    // there is no inline version of llvm.memmove)
+    if (isa<MemCpyInlineInst>(M))
+      return false;
     UseMemMove = true;
+  }
 
   // If all checks passed, then we can transform M.
   LLVM_DEBUG(dbgs() << "MemCpyOptPass: Forwarding memcpy->memcpy src:\n"
diff --git a/llvm/test/Transforms/MemCpyOpt/memcpy.ll b/llvm/test/Transforms/MemCpyOpt/memcpy.ll
--- a/llvm/test/Transforms/MemCpyOpt/memcpy.ll
+++ b/llvm/test/Transforms/MemCpyOpt/memcpy.ll
@@ -125,6 +125,20 @@
 
 }
 
+; Similar to test5_memcpy, but without noalias; check that memcpy.inline is not folded into memmove.
+define void @test6_memcpy(ptr %src, ptr %dest) nounwind {
+; CHECK-LABEL: @test6_memcpy(
+; CHECK-NEXT:    [[TMP:%.*]] = alloca [16 x i8], align 1
+; CHECK-NEXT:    call void @llvm.memcpy.inline.p0.p0.i32(ptr align 1 [[TMP]], ptr align 1 [[DEST:%.*]], i32 16, i1 false)
+; CHECK-NEXT:    call void @llvm.memcpy.inline.p0.p0.i32(ptr align 1 [[DEST]], ptr align 1 [[TMP]], i32 16, i1 false)
+; CHECK-NEXT:    ret void
+;
+  %tmp = alloca [16 x i8], align 1
+  call void @llvm.memcpy.inline.p0.p0.i32(ptr align 1 %tmp, ptr align 1 %dest, i32 16, i1 false)
+  call void @llvm.memcpy.inline.p0.p0.i32(ptr align 1 %dest, ptr align 1 %tmp, i32 16, i1 false)
+  ret void
+}
+
 
 @x = external global %0