Index: include/llvm/Analysis/TargetTransformInfoImpl.h
===================================================================
--- include/llvm/Analysis/TargetTransformInfoImpl.h
+++ include/llvm/Analysis/TargetTransformInfoImpl.h
@@ -785,13 +785,24 @@
     if (getUserCost(I, Operands) == TTI::TCC_Free)
       return 0;
 
-    if (isa<CallInst>(I))
-      return 40;
-
     if (isa<LoadInst>(I))
       return 4;
 
     Type *dstTy = I->getType();
+
+    // Usually an intrinsic is a simple instruction.
+    // A real function call is much slower.
+    if (auto *CI = dyn_cast<CallInst>(I)) {
+      const Function *F = CI->getCalledFunction();
+      if (static_cast<T *>(this)->isLoweredToCall(F))
+        return 40;
+      // Some intrinsics return a value and a flag, we use the value type
+      // to decide its latency.
+      if (StructType* StructTy = dyn_cast<StructType>(dstTy))
+        dstTy = StructTy->getElementType(0);
+      // Fall through to simple instructions.
+    }
+
     if (VectorType *VectorTy = dyn_cast<VectorType>(dstTy))
       dstTy = VectorTy->getElementType();
     if (dstTy->isFloatingPointTy())
Index: test/Analysis/CostModel/X86/costmodel.ll
===================================================================
--- test/Analysis/CostModel/X86/costmodel.ll
+++ test/Analysis/CostModel/X86/costmodel.ll
@@ -5,6 +5,8 @@
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 
+declare { i32, i1 } @llvm.uadd.with.overflow.i32(i32, i32)
+
 define i64 @foo(i64 %arg) {
 
   ; LATENCY:  cost of 0 {{.*}} alloca i32
@@ -39,6 +41,10 @@
   ; CODESIZE: cost of 0 {{.*}} trunc
   %TC = trunc i64 undef to i32
 
+  ; LATENCY:  cost of 1 {{.*}} call
+  ; CODESIZE: cost of 1 {{.*}} call
+  %uadd = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 undef, i32 undef)
+
   ; LATENCY:  cost of 1 {{.*}} ret
   ; CODESIZE: cost of 1 {{.*}} ret
   ret i64 undef