[LTO/gold] Support --wrap

[tejohnson]

Summary

Utilize new gold plugin api interface for obtaining --wrap option
arguments, and LTO API handling (added for --wrap support in lld LTO),
to mark symbols so that LTO does not optimize them inappropriately.

Note the test cases will be in a new gold test subdirectory that
is dependent on the next release of gold which will contain the new
interfaces.

[tejohnson]

ping

[pcc]

LGTM

tools/gold/gold-plugin.cpp

594

Seems a little odd to have this loop go in reverse. Unless there is some reason to do that, I'd write it in the conventional way.

[tejohnson]

Address comments

[tejohnson]

I had to revert as the new api is not available:
http://lab.llvm.org:8011/builders/clang-with-thin-lto-ubuntu/builds/9109/steps/build-stage2-LLVMgold.so/logs/stdio

What is the best way to handle this? I see that there are a couple defines at the top of gold-plugin.cpp for other newer LDPT_* types, so I could add LDPT_GET_WRAP_SYMBOLS there.

For ld_plugin_get_wrap_symbols I guess I can add my own typedef at the top of gold-plugin.cpp?

[pcc]

In D44235#1036549, @tejohnson wrote:

I had to revert as the new api is not available:
http://lab.llvm.org:8011/builders/clang-with-thin-lto-ubuntu/builds/9109/steps/build-stage2-LLVMgold.so/logs/stdio

What is the best way to handle this? I see that there are a couple defines at the top of gold-plugin.cpp for other newer LDPT_* types, so I could add LDPT_GET_WRAP_SYMBOLS there.

For ld_plugin_get_wrap_symbols I guess I can add my own typedef at the top of gold-plugin.cpp?

Sounds good. Of course, the tv_get_wrap_symbols field will not be present so you will need to cast one of the other fields.

[tejohnson]

In D44235#1036553, @pcc wrote:

In D44235#1036549, @tejohnson wrote:

I had to revert as the new api is not available:
http://lab.llvm.org:8011/builders/clang-with-thin-lto-ubuntu/builds/9109/steps/build-stage2-LLVMgold.so/logs/stdio

What is the best way to handle this? I see that there are a couple defines at the top of gold-plugin.cpp for other newer LDPT_* types, so I could add LDPT_GET_WRAP_SYMBOLS there.

For ld_plugin_get_wrap_symbols I guess I can add my own typedef at the top of gold-plugin.cpp?

Sounds good. Of course, the tv_get_wrap_symbols field will not be present so you will need to cast one of the other fields.

Ugh, yeah that too. It's a little ugly, but I added FIXME comments in the various places to remove them when we require the binutils 2.31 (which would be the next version). I tested with both the old and new plugin-api.h to make sure it works with both now. I'll submit in the morning so that I can watch the bots.

[tejohnson]

Workarounds for old plugin-api.h

[tejohnson]

Fix binutils version in FIXME

[pcc]

LGTM

test/tools/gold/X86/v1.16/Inputs/wrap-bar.ll

+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define hidden void @bar() {
+  ret void
+}
+
+define hidden void @__real_bar() {
+  ret void
+}
+
+define hidden void @__wrap_bar() {
+  ret void
+}

test/tools/gold/X86/v1.16/lit.local.cfg

+import re
+import subprocess
+
+def is_gold_v1_16_linker_available():
+
+  if not config.gold_executable:
+    return False
+  try:
+    ld_cmd = subprocess.Popen([config.gold_executable, '-v'],
+                              stdout = subprocess.PIPE,
+                              stderr = subprocess.PIPE)
+    ld_out, _ = ld_cmd.communicate()
+    ld_out = ld_out.decode()
+  except:
+    return False
+
+  match = re.search(r'GNU gold \(.*\) (\d+)\.(\d+)', ld_out)
+  if not match:
+    return False
+  major = int(match.group(1))
+  minor = int(match.group(2))
+  if major < 1 or (major == 1 and minor < 16):
+    return False
+
+  return True
+
+if not is_gold_v1_16_linker_available():
+  config.unsupported = True

test/tools/gold/X86/v1.16/wrap-1.ll

+; LTO
+; RUN: llvm-as %s -o %t.o
+; RUN: %gold -plugin %llvmshlibdir/LLVMgold%shlibext %t.o -o %t.out -wrap=bar -plugin-opt=save-temps
+; RUN: llvm-readobj -t %t.out | FileCheck %s
+; RUN: cat %t.out.resolution.txt | FileCheck -check-prefix=RESOLS %s
+
+; ThinLTO
+; RUN: opt -module-summary %s -o %t.o
+; RUN: %gold -plugin %llvmshlibdir/LLVMgold%shlibext %t.o -o %t.out -wrap=bar -plugin-opt=save-temps
+; RUN: llvm-readobj -t %t.out | FileCheck %s
+; RUN: cat %t.out.resolution.txt | FileCheck -check-prefix=RESOLS %s
+
+; CHECK:      Name: __wrap_bar
+; CHECK-NEXT: Value:
+; CHECK-NEXT: Size:
+; CHECK-NEXT: Binding: Global
+; CHECK-NEXT: Type: Function
+
+; Make sure that the 'r' (linker redefined) bit is set for bar and __real_bar
+; in the resolutions file, and that the 'x' (visible to regular obj) bit is set
+; for bar and __wrap_bar.
+; RESOLS: ,bar,lxr
+; RESOLS: ,__wrap_bar,plx
+; RESOLS: ,__real_bar,plr
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+declare void @bar()
+
+define void @_start() {
+  call void @bar()
+  ret void
+}
+
+define void @__wrap_bar() {
+  ret void
+}
+
+define void @__real_bar() {
+  ret void
+}

test/tools/gold/X86/v1.16/wrap-2.ll

+; LTO
+; This doesn't currently work with gold, because it does not apply defsym
+; renaming to symbols in the same module (apparently by design for consistency
+; with GNU ld). Because regular LTO hands back a single object file to gold,
+; it doesn't perform the desired defsym renaming. This isn't an issue with
+; ThinLTO which hands back multiple native objects to gold. For regular
+; LTO defsym handling, gold will need a fix (not the gold plugin).
+; RUN-TODO: llvm-as %s -o %t.o
+; RUN-TODO: llvm-as %S/Inputs/wrap-bar.ll -o %t1.o
+; RUN-TODO: %gold -plugin %llvmshlibdir/LLVMgold%shlibext %t.o %t1.o -shared -o %t.so -wrap=bar
+; RUN-TODO: llvm-objdump -d %t.so | FileCheck %s
+; RUN-TODO: llvm-readobj -t %t.so | FileCheck -check-prefix=BIND %s
+
+; ThinLTO
+; RUN: opt -module-summary %s -o %t.o
+; RUN: opt -module-summary %S/Inputs/wrap-bar.ll -o %t1.o
+; RUN: %gold -plugin %llvmshlibdir/LLVMgold%shlibext %t.o %t1.o -shared -o %t.so -wrap=bar
+; RUN: llvm-objdump -d %t.so | FileCheck %s -check-prefix=THIN
+; RUN: llvm-readobj -t %t.so | FileCheck -check-prefix=BIND %s
+
+; Make sure that calls in foo() are not eliminated and that bar is
+; routed to __wrap_bar and __real_bar is routed to bar.
+
+; CHECK:      foo:
+; CHECK-NEXT: pushq %rax
+; CHECK-NEXT: callq{{.*}}<__wrap_bar>
+; CHECK-NEXT: callq{{.*}}<bar>
+
+; THIN:      foo:
+; THIN-NEXT: pushq  %rax
+; THIN-NEXT: callq{{.*}}<__wrap_bar>
+; THIN-NEXT: popq  %rax
+; THIN-NEXT: jmp{{.*}}<bar>
+
+; Check that bar and __wrap_bar retain their original binding.
+; BIND:      Name: bar
+; BIND-NEXT: Value:
+; BIND-NEXT: Size:
+; BIND-NEXT: Binding: Local
+; BIND:      Name: __wrap_bar
+; BIND-NEXT: Value:
+; BIND-NEXT: Size:
+; BIND-NEXT: Binding: Local
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+declare void @bar()
+declare void @__real_bar()
+
+define void @foo() {
+  call void @bar()
+  call void @__real_bar()
+  ret void
+}

tools/gold/gold-plugin.cpp

 #include <vector>
 
 // FIXME: remove this declaration when we stop maintaining Ubuntu Quantal and
 // Precise and Debian Wheezy (binutils 2.23 is required)
 #define LDPO_PIE 3
 
 #define LDPT_GET_SYMBOLS_V3 28
 
+// FIXME: Remove when binutils 2.31 (containing gold 1.16) is the minimum
+// required version.
+#define LDPT_GET_WRAP_SYMBOLS 32
+
 using namespace llvm;
 using namespace lto;
 
+// FIXME: Remove when binutils 2.31 (containing gold 1.16) is the minimum
+// required version.
+typedef enum ld_plugin_status (*ld_plugin_get_wrap_symbols)(
+    uint64_t *num_symbols, const char ***wrap_symbol_list);
+
 static ld_plugin_status discard_message(int level, const char *format, ...) {
   // Die loudly. Recent versions of Gold pass ld_plugin_message as the first
   // callback in the transfer vector. This should never be called.
   abort();
 }
 
 static ld_plugin_release_input_file release_input_file = nullptr;
 static ld_plugin_get_input_file get_input_file = nullptr;
 static ld_plugin_message message = discard_message;
+static ld_plugin_get_wrap_symbols get_wrap_symbols = nullptr;
 
 namespace {
 struct claimed_file {
   void *handle;
   void *leader_handle;
   std::vector<ld_plugin_symbol> syms;
   off_t filesize;
   std::string name;
struct PluginInputFile {
 
   PluginInputFile(PluginInputFile &&RHS) = default;
   PluginInputFile &operator=(PluginInputFile &&RHS) = default;
 };
 
 struct ResolutionInfo {
   bool CanOmitFromDynSym = true;
   bool DefaultVisibility = true;
+  bool CanInline = true;
+  bool IsUsedInRegularObj = false;
 };
 
 }
 
 static ld_plugin_add_symbols add_symbols = nullptr;
 static ld_plugin_get_symbols get_symbols = nullptr;
 static ld_plugin_add_input_file add_input_file = nullptr;
 static ld_plugin_set_extra_library_path set_extra_library_path = nullptr;
case LDPT_SET_EXTRA_LIBRARY_PATH:
       set_extra_library_path = tv->tv_u.tv_set_extra_library_path;
       break;
     case LDPT_GET_VIEW:
       get_view = tv->tv_u.tv_get_view;
       break;
     case LDPT_MESSAGE:
       message = tv->tv_u.tv_message;
       break;
+    case LDPT_GET_WRAP_SYMBOLS:
+      // FIXME: When binutils 2.31 (containing gold 1.16) is the minimum
+      // required version, this should be changed to:
+      // get_wrap_symbols = tv->tv_u.tv_get_wrap_symbols;
+      get_wrap_symbols =
+          (ld_plugin_get_wrap_symbols)tv->tv_u.tv_register_new_input;
+      break;
     default:
       break;
     }
   }
 
   if (!registeredClaimFile) {
     message(LDPL_ERROR, "register_claim_file not passed to LLVMgold.");
     return LDPS_ERR;
static ld_plugin_status claim_file_hook(const ld_plugin_input_file *file,
 
   if (!cf.syms.empty()) {
     if (add_symbols(cf.handle, cf.syms.size(), cf.syms.data()) != LDPS_OK) {
       message(LDPL_ERROR, "Unable to add symbols!");
       return LDPS_ERR;
     }
   }
 
+  // Handle any --wrap options passed to gold, which are than passed
+  // along to the plugin.
+  if (get_wrap_symbols) {
+    const char **wrap_symbols;
+    uint64_t count = 0;
+    if (get_wrap_symbols(&count, &wrap_symbols) != LDPS_OK) {
+      message(LDPL_ERROR, "Unable to get wrap symbols!");
+      return LDPS_ERR;
+    }
+    for (uint64_t i = 0; i < count; i++) {
+      StringRef Name = wrap_symbols[i];
+      ResolutionInfo &Res = ResInfo[Name];
+      ResolutionInfo &WrapRes = ResInfo["__wrap_" + Name.str()];
+      ResolutionInfo &RealRes = ResInfo["__real_" + Name.str()];
+      // Tell LTO not to inline symbols that will be overwritten.
+      Res.CanInline = false;
+      RealRes.CanInline = false;
+      // Tell LTO not to eliminate symbols that will be used after renaming.
+      Res.IsUsedInRegularObj = true;
+      WrapRes.IsUsedInRegularObj = true;
+    }
+  }
+
   return LDPS_OK;
 }
 
 static void freeSymName(ld_plugin_symbol &Sym) {
   free(Sym.name);
   free(Sym.comdat_key);
   Sym.name = nullptr;
   Sym.comdat_key = nullptr;
for (ld_plugin_symbol &Sym : F.syms) {
     // __stop_<secname> symbols which may be used elsewhere.
     if (isValidCIdentifier(InpSym.getSectionName()))
       R.VisibleToRegularObj = true;
 
     if (Resolution != LDPR_RESOLVED_DYN && Resolution != LDPR_UNDEF &&
         (IsExecutable || !Res.DefaultVisibility))
       R.FinalDefinitionInLinkageUnit = true;
 
+    if (!Res.CanInline)
+      R.LinkerRedefined = true;
+
+    if (Res.IsUsedInRegularObj)
+      R.VisibleToRegularObj = true;
+
     freeSymName(Sym);
   }
 
   check(Lto.add(std::move(Input), Resols),
         std::string("Failed to link module ") + F.name);
 }
 
 static void recordFile(const std::string &Filename, bool TempOutFile) {