diff --git a/clang/lib/Driver/Driver.cpp b/clang/lib/Driver/Driver.cpp
--- a/clang/lib/Driver/Driver.cpp
+++ b/clang/lib/Driver/Driver.cpp
@@ -2024,7 +2024,8 @@
 
     // Lipo if necessary, we do it this way because we need to set the arch flag
     // so that -Xarch_ gets overwritten.
-    if (Inputs.size() == 1 || Act->getType() == types::TY_Nothing)
+    bool LipoNeeded = Inputs.size() != 1 && Act->getType() != types::TY_Nothing;
+    if (!LipoNeeded)
       Actions.append(Inputs.begin(), Inputs.end());
     else
       Actions.push_back(C.MakeAction<LipoJobAction>(Inputs, Act->getType()));
@@ -2034,16 +2035,17 @@
     bool enablesDebugInfo = A && !A->getOption().matches(options::OPT_g0) &&
                             !A->getOption().matches(options::OPT_gstabs);
     bool isGeneratingTemporaryObject =
-        ContainsCompileOrAssembleAction(Actions.back()) || isUsingLTO();
+        ContainsCompileOrAssembleAction(Actions.back()) || isUsingLTO() ||
+        LipoNeeded;
     if ((enablesDebugInfo || willEmitRemarks(Args)) &&
         isGeneratingTemporaryObject) {
 
       // Add a 'dsymutil' step if necessary, when debug info is enabled and we
       // are linking a temporary object. This occurs when we have a compiler
-      // or assmbler input or if LTO is enabled. We need to run 'dsymutil'
-      // ourselves in such cases because the debug info will refer to the
-      // temporary object file which will be removed at the end of the
-      // compilation process.
+      // or assembler input, if LTO is enabled, or if we're building multiple
+      // architectures. We need to run 'dsymutil' ourselves in such cases
+      // because the debug info will refer to the temporary object file which
+      // will be removed at the end of the compilation process.
       if (Act->getType() == types::TY_Image) {
         ActionList Inputs;
         Inputs.push_back(Actions.back());
diff --git a/clang/lib/Driver/ToolChains/Darwin.h b/clang/lib/Driver/ToolChains/Darwin.h
--- a/clang/lib/Driver/ToolChains/Darwin.h
+++ b/clang/lib/Driver/ToolChains/Darwin.h
@@ -59,10 +59,10 @@
 };
 
 class LLVM_LIBRARY_VISIBILITY Linker : public MachOTool {
-  bool NeedsTempPath(const InputInfoList &Inputs,
+  bool NeedsTempPath(const InputInfo &Outputs, const InputInfoList &Inputs,
                      const llvm::opt::ArgList &Args) const;
   void AddLinkArgs(Compilation &C, const llvm::opt::ArgList &Args,
-                   llvm::opt::ArgStringList &CmdArgs,
+                   llvm::opt::ArgStringList &CmdArgs, const InputInfo &Outputs,
                    const InputInfoList &Inputs, unsigned Version[5]) const;
 
 public:
diff --git a/clang/lib/Driver/ToolChains/Darwin.cpp b/clang/lib/Driver/ToolChains/Darwin.cpp
--- a/clang/lib/Driver/ToolChains/Darwin.cpp
+++ b/clang/lib/Driver/ToolChains/Darwin.cpp
@@ -167,7 +167,8 @@
     CmdArgs.push_back("-force_cpusubtype_ALL");
 }
 
-bool darwin::Linker::NeedsTempPath(const InputInfoList &Inputs,
+bool darwin::Linker::NeedsTempPath(const InputInfo &Output,
+                                   const InputInfoList &Inputs,
                                    const ArgList &Args) const {
   Arg *A = Args.getLastArg(options::OPT_g_Group);
   if (A && !A->getOption().matches(options::OPT_g0) &&
@@ -181,6 +182,11 @@
     if (Input.getType() != types::TY_Object)
       return true;
 
+  // If we're going to use lipo (i.e. the linker output is a temporary file too)
+  // then we'll need to run dsymutil after linking too
+  if (Inputs.size() != 1 && Output.getType() != types::TY_Nothing)
+    return true;
+
   return false;
 }
 
@@ -209,6 +215,7 @@
 
 void darwin::Linker::AddLinkArgs(Compilation &C, const ArgList &Args,
                                  ArgStringList &CmdArgs,
+                                 const InputInfo &Output,
                                  const InputInfoList &Inputs,
                                  unsigned Version[5]) const {
   const Driver &D = getToolChain().getDriver();
@@ -228,7 +235,8 @@
                    options::OPT_fno_application_extension, false))
     CmdArgs.push_back("-application_extension");
 
-  if (D.isUsingLTO() && Version[0] >= 116 && NeedsTempPath(Inputs, Args)) {
+  if (D.isUsingLTO() && Version[0] >= 116 &&
+      NeedsTempPath(Output, Inputs, Args)) {
     std::string TmpPathName;
     if (D.getLTOMode() == LTOK_Full) {
       // If we are using full LTO, then automatically create a temporary file
@@ -541,7 +549,7 @@
 
   // I'm not sure why this particular decomposition exists in gcc, but
   // we follow suite for ease of comparison.
-  AddLinkArgs(C, Args, CmdArgs, Inputs, Version);
+  AddLinkArgs(C, Args, CmdArgs, Output, Inputs, Version);
 
   if (willEmitRemarks(Args) &&
       checkRemarksOptions(getToolChain().getDriver(), Args,
diff --git a/clang/test/Driver/darwin-dsymutil.c b/clang/test/Driver/darwin-dsymutil.c
--- a/clang/test/Driver/darwin-dsymutil.c
+++ b/clang/test/Driver/darwin-dsymutil.c
@@ -61,3 +61,14 @@
 
 // CHECK-DSYMUTIL-LTO: "/usr/bin/ld" "-demangle" "-object_path_lto"
 // CHECK-DSYMUTIL-LTO: "/usr/bin/dsymutil" "-o" "a.out.dSYM" "a.out"
+
+// Check that clang provides the linker with a temporary object file path and
+// runs dsymutil when building for multiple architectures.
+// RUN: touch %t.o
+// RUN: %clang -target x86_64-apple-darwin10 -arch x86_64 -arch arm64 %t.o -g \
+// RUN:     -flto -### 2> %t
+// RUN: FileCheck -check-prefix=CHECK-DSYMUTIL-MULTIARCH < %t %s
+
+// CHECK-DSYMUTIL-MULTIARCH: "/usr/bin/ld" "-demangle" "-object_path_lto"
+// CHECK-DSYMUTIL-MULTIARCH: "/usr/bin/ld" "-demangle" "-object_path_lto"
+// CHECK-DSYMUTIL-MULTIARCH: "/usr/bin/dsymutil" "-o" "a.out.dSYM" "a.out"