diff --git a/lib/Target/X86/x86_64Relocator.cpp b/lib/Target/X86/x86_64Relocator.cpp
index 3af4b4d6f..990416bd4 100644
--- a/lib/Target/X86/x86_64Relocator.cpp
+++ b/lib/Target/X86/x86_64Relocator.cpp
@@ -274,9 +274,17 @@ void x86_64Relocator::scanLocalReloc(InputFile &pInputFile, Relocation &pReloc,
     std::lock_guard<std::mutex> relocGuard(m_RelocMutex);
     if (rsym->reserved() & ReserveGOT)
       return;
-    // Create a TLS IE GOT entry.
     x86_64GOT *G = m_Target.createGOT(GOT::TLS_IE, Obj, rsym);
-    G->setValueType(GOT::TLSStaticSymbolValue);
+    // For executables, the symbol's offset from the thread pointer is fixed at
+    // link time. For shared objects, the dynamic loader must compute the offset
+    // at load time, so emit R_X86_64_TPOFF64.
+    if (config().isBuildingExecutable()) {
+      G->setValueType(GOT::TLSStaticSymbolValue);
+    } else {
+      helper_DynRel_init(Obj, &pReloc, rsym, G, 0x0,
+                         llvm::ELF::R_X86_64_TPOFF64, m_Target);
+      m_Target.setHasStaticTLS();
+    }
     rsym->setReserved(rsym->reserved() | ReserveGOT);
     return;
   }
@@ -408,13 +416,14 @@ void x86_64Relocator::scanGlobalReloc(InputFile &pInputFile, Relocation &pReloc,
     if (rsym->reserved() & ReserveGOT)
       return;
     x86_64GOT *G = m_Target.createGOT(GOT::TLS_IE, Obj, rsym);
-    const bool isExec = (config().codeGenType() == LinkerConfig::Exec);
+    const bool isExec = config().isBuildingExecutable();
     const bool preemptible = m_Target.isSymbolPreemptible(*rsym);
     if (isExec && !preemptible) {
       G->setValueType(GOT::TLSStaticSymbolValue);
     } else {
       helper_DynRel_init(Obj, &pReloc, rsym, G, 0x0,
                          llvm::ELF::R_X86_64_TPOFF64, m_Target);
+      m_Target.setHasStaticTLS();
     }
     rsym->setReserved(rsym->reserved() | ReserveGOT);
     return;
diff --git a/test/x86_64/linux/TLSIEHiddenGlobal/TLSIEHiddenGlobal.test b/test/x86_64/linux/TLSIEHiddenGlobal/TLSIEHiddenGlobal.test
new file mode 100644
index 000000000..31ea691ba
--- /dev/null
+++ b/test/x86_64/linux/TLSIEHiddenGlobal/TLSIEHiddenGlobal.test
@@ -0,0 +1,19 @@
+/// Verifies the global-symbol path of R_X86_64_GOTTPOFF for a non-preemptible
+/// TLS variable. In -pie the symbol's TP offset is fixed at link time,
+/// so no R_X86_64_TPOFF64 should be emitted. In -shared the
+/// dynamic loader resolves the offset at load time, so the dyn reloc and
+/// DF_STATIC_TLS are both required.
+
+// RUN: %clang %clangopts -c -x c %s -fPIC -ftls-model=initial-exec -o %t.o
+// RUN: %link %linkopts -shared %t.o -o %t.shared.so
+// RUN: %readelf -r -d %t.shared.so | %filecheck %s --check-prefix=SHARED
+// RUN: %link %linkopts -pie %t.o -o %t.pie --defsym __libc_start_main=0
+// RUN: %readelf -r %t.pie | %filecheck %s --check-prefix=PIE
+
+// SHARED: STATIC_TLS
+// SHARED: R_X86_64_TPOFF64
+
+// PIE-NOT: R_X86_64_TPOFF64
+
+__attribute__((visibility("hidden"))) __thread int g = 20;
+int bar(void) { return g; }
diff --git a/test/x86_64/linux/TLSIELocalSymbol/TLSIELocalSymbol.test b/test/x86_64/linux/TLSIELocalSymbol/TLSIELocalSymbol.test
new file mode 100644
index 000000000..07e856414
--- /dev/null
+++ b/test/x86_64/linux/TLSIELocalSymbol/TLSIELocalSymbol.test
@@ -0,0 +1,21 @@
+/// Verifies that R_X86_64_GOTTPOFF against a file-local TLS symbol emits a
+/// R_X86_64_TPOFF64 dynamic relocation in -shared output, and is statically
+/// resolved into the GOT slot in -pie.
+
+// RUN: %clang %clangopts -c -x c %s -fPIC -ftls-model=initial-exec -o %t.o
+// RUN: %link %linkopts -shared %t.o -o %t.shared.so
+// RUN: %readelf -r -x .got -d %t.shared.so | %filecheck %s --check-prefix=SHARED
+// RUN: %link %linkopts -pie %t.o -o %t.pie --defsym __libc_start_main=0
+// RUN: %readelf -r -x .got %t.pie | %filecheck %s --check-prefix=PIE
+
+// SHARED: STATIC_TLS
+// SHARED: R_X86_64_TPOFF64
+// SHARED-LABEL: Hex dump of section '.got':
+// SHARED-NEXT: 0x{{[0-9a-f]+}} 00000000 00000000
+
+// PIE-NOT: R_X86_64_TPOFF64
+// PIE-LABEL: Hex dump of section '.got':
+// PIE-NEXT: 0x{{[0-9a-f]+}} fcffffff ffffffff
+
+static __thread int a = 10;
+int foo(void) { return a; }