Index: lib/LTO/LTOCodeGenerator.cpp
===================================================================
--- lib/LTO/LTOCodeGenerator.cpp
+++ lib/LTO/LTOCodeGenerator.cpp
@@ -370,8 +370,8 @@
 
 static void accumulateAndSortLibcalls(std::vector<StringRef> &Libcalls,
                                       const TargetLibraryInfo& TLI,
-                                      const TargetLowering *Lowering)
-{
+                                      const Module &Mod,
+                                      const TargetMachine &TM) {
   // TargetLibraryInfo has info on C runtime library calls on the current
   // target.
   for (unsigned I = 0, E = static_cast<unsigned>(LibFunc::NumLibFuncs);
@@ -381,14 +381,21 @@
       Libcalls.push_back(TLI.getName(F));
   }
 
-  // TargetLowering has info on library calls that CodeGen expects to be
-  // available, both from the C runtime and compiler-rt.
-  if (Lowering)
-    for (unsigned I = 0, E = static_cast<unsigned>(RTLIB::UNKNOWN_LIBCALL);
-         I != E; ++I)
-      if (const char *Name
-          = Lowering->getLibcallName(static_cast<RTLIB::Libcall>(I)))
-        Libcalls.push_back(Name);
+  SmallPtrSet<const TargetLowering *, 1> TLSet;
+
+  for (const Function &F : Mod) {
+    const TargetLowering *Lowering =
+        TM.getSubtargetImpl(F)->getTargetLowering();
+
+    if (Lowering && TLSet.insert(Lowering).second)
+      // TargetLowering has info on library calls that CodeGen expects to be
+      // available, both from the C runtime and compiler-rt.
+      for (unsigned I = 0, E = static_cast<unsigned>(RTLIB::UNKNOWN_LIBCALL);
+           I != E; ++I)
+        if (const char *Name =
+                Lowering->getLibcallName(static_cast<RTLIB::Libcall>(I)))
+          Libcalls.push_back(Name);
+  }
 
   array_pod_sort(Libcalls.begin(), Libcalls.end());
   Libcalls.erase(std::unique(Libcalls.begin(), Libcalls.end()),
@@ -412,8 +419,8 @@
   std::vector<StringRef> Libcalls;
   TargetLibraryInfoImpl TLII(Triple(TargetMach->getTargetTriple()));
   TargetLibraryInfo TLI(TLII);
-  accumulateAndSortLibcalls(
-      Libcalls, TLI, TargetMach->getSubtargetImpl()->getTargetLowering());
+
+  accumulateAndSortLibcalls(Libcalls, TLI, *mergedModule, *TargetMach);
 
   for (Module::iterator f = mergedModule->begin(),
          e = mergedModule->end(); f != e; ++f)
Index: test/LTO/runtime-library-subtarget.ll
===================================================================
--- /dev/null
+++ test/LTO/runtime-library-subtarget.ll
@@ -0,0 +1,18 @@
+; Check that user-defined runtime library function __addsf3vfp is not removed
+;
+; RUN: llvm-as <%s >%t1
+; RUN: llvm-lto -o %t2 %t1 -mcpu arm1176jz-s
+; RUN: llvm-nm %t2 | FileCheck %s
+
+target datalayout = "e-m:o-p:32:32-f64:32:64-v64:32:64-v128:32:128-a:0:32-n32-S32"
+target triple = "thumbv7-apple-ios"
+
+; CHECK: ___addsf3vfp
+
+define float @__addsf3vfp(float %a, float %b) #0 {
+entry:
+  %add = fadd float %a, %b
+  ret float %add
+}
+
+attributes #0 = { "target-cpu"="arm1176jzf-s"}
Index: tools/llvm-lto/llvm-lto.cpp
===================================================================
--- tools/llvm-lto/llvm-lto.cpp
+++ tools/llvm-lto/llvm-lto.cpp
@@ -217,6 +217,8 @@
   for (unsigned i = 0; i < KeptDSOSyms.size(); ++i)
     CodeGen.addMustPreserveSymbol(KeptDSOSyms[i].c_str());
 
+  CodeGen.setCpu(MCPU.c_str());
+
   std::string attrs;
   for (unsigned i = 0; i < MAttrs.size(); ++i) {
     if (i > 0)