[llvm][GlobalOpt] Remove empty atexit destructors/handlers (#88836)

https://godbolt.org/z/frjhqMKqc for an example.

Removal of allocations due to empty `__cxa_atexit` destructor calls is
done by the following globalopt pass.
This pass currently does not look for `atexit` handlers generated for
platforms that do not use `__cxa_atexit`.
By default Win32 and AIX use `atexit`.

I don't see an easy way to only remove `atexit` calls that the compiler
generated without looking at the generated mangled name of the atexit
handler that is being registered.
However we can easily remove all `atexit` calls that register empty
handlers since it is trivial to ensure the removed call still returns
`0` which is the value for success.

Author: MaxEW707

llvm/include/llvm/Analysis/TargetLibraryInfo.def

@@ -471,6 +471,11 @@ TLI_DEFINE_ENUM_INTERNAL(cxa_atexit)
 TLI_DEFINE_STRING_INTERNAL("__cxa_atexit")
 TLI_DEFINE_SIG_INTERNAL(Int, Ptr, Ptr, Ptr)
 
+/// int atexit(void (*f)(void));
+TLI_DEFINE_ENUM_INTERNAL(atexit)
+TLI_DEFINE_STRING_INTERNAL("atexit")
+TLI_DEFINE_SIG_INTERNAL(Int, Ptr)
+
 /// void __cxa_guard_abort(guard_t *guard);
 /// guard_t is int64_t in Itanium ABI or int32_t on ARM eabi.
 TLI_DEFINE_ENUM_INTERNAL(cxa_guard_abort)

llvm/lib/Transforms/IPO/GlobalOpt.cpp

@@ -87,6 +87,7 @@ STATISTIC(NumNestRemoved   , "Number of nest attributes removed");
 STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
 STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed");
+STATISTIC(NumAtExitRemoved, "Number of atexit handlers removed");
 STATISTIC(NumInternalFunc, "Number of internal functions");
 STATISTIC(NumColdCC, "Number of functions marked coldcc");
 STATISTIC(NumIFuncsResolved, "Number of statically resolved IFuncs");
@@ -2328,36 +2329,38 @@ OptimizeGlobalAliases(Module &M,
 }
 
 static Function *
-FindCXAAtExit(Module &M, function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
+FindAtExitLibFunc(Module &M,
+                  function_ref<TargetLibraryInfo &(Function &)> GetTLI,
+                  LibFunc Func) {
   // Hack to get a default TLI before we have actual Function.
   auto FuncIter = M.begin();
   if (FuncIter == M.end())
     return nullptr;
   auto *TLI = &GetTLI(*FuncIter);
 
-  LibFunc F = LibFunc_cxa_atexit;
-  if (!TLI->has(F))
+  if (!TLI->has(Func))
     return nullptr;
 
-  Function *Fn = M.getFunction(TLI->getName(F));
+  Function *Fn = M.getFunction(TLI->getName(Func));
   if (!Fn)
     return nullptr;
 
   // Now get the actual TLI for Fn.
   TLI = &GetTLI(*Fn);
 
   // Make sure that the function has the correct prototype.
-  if (!TLI->getLibFunc(*Fn, F) || F != LibFunc_cxa_atexit)
+  LibFunc F;
+  if (!TLI->getLibFunc(*Fn, F) || F != Func)
     return nullptr;
 
   return Fn;
 }
 
-/// Returns whether the given function is an empty C++ destructor and can
-/// therefore be eliminated.
-/// Note that we assume that other optimization passes have already simplified
-/// the code so we simply check for 'ret'.
-static bool cxxDtorIsEmpty(const Function &Fn) {
+/// Returns whether the given function is an empty C++ destructor or atexit
+/// handler and can therefore be eliminated. Note that we assume that other
+/// optimization passes have already simplified the code so we simply check for
+/// 'ret'.
+static bool IsEmptyAtExitFunction(const Function &Fn) {
   // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
   // nounwind, but that doesn't seem worth doing.
   if (Fn.isDeclaration())
@@ -2373,7 +2376,7 @@ static bool cxxDtorIsEmpty(const Function &Fn) {
   return false;
 }
 
-static bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
+static bool OptimizeEmptyGlobalAtExitDtors(Function *CXAAtExitFn, bool isCXX) {
   /// Itanium C++ ABI p3.3.5:
   ///
   ///   After constructing a global (or local static) object, that will require
@@ -2386,8 +2389,8 @@ static bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
   ///   registered before this one. It returns zero if registration is
   ///   successful, nonzero on failure.
 
-  // This pass will look for calls to __cxa_atexit where the function is trivial
-  // and remove them.
+  // This pass will look for calls to __cxa_atexit or atexit where the function
+  // is trivial and remove them.
   bool Changed = false;
 
   for (User *U : llvm::make_early_inc_range(CXAAtExitFn->users())) {
@@ -2400,14 +2403,17 @@ static bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
 
     Function *DtorFn =
       dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
-    if (!DtorFn || !cxxDtorIsEmpty(*DtorFn))
+    if (!DtorFn || !IsEmptyAtExitFunction(*DtorFn))
       continue;
 
     // Just remove the call.
     CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
     CI->eraseFromParent();
 
-    ++NumCXXDtorsRemoved;
+    if (isCXX)
+      ++NumCXXDtorsRemoved;
+    else
+      ++NumAtExitRemoved;
 
     Changed |= true;
   }
@@ -2525,9 +2531,12 @@ optimizeGlobalsInModule(Module &M, const DataLayout &DL,
 
     // Try to remove trivial global destructors if they are not removed
     // already.
-    Function *CXAAtExitFn = FindCXAAtExit(M, GetTLI);
-    if (CXAAtExitFn)
-      LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
+    if (Function *CXAAtExitFn =
+            FindAtExitLibFunc(M, GetTLI, LibFunc_cxa_atexit))
+      LocalChange |= OptimizeEmptyGlobalAtExitDtors(CXAAtExitFn, true);
+
+    if (Function *AtExitFn = FindAtExitLibFunc(M, GetTLI, LibFunc_atexit))
+      LocalChange |= OptimizeEmptyGlobalAtExitDtors(AtExitFn, false);
 
     // Optimize IFuncs whose callee's are statically known.
     LocalChange |= OptimizeStaticIFuncs(M);

llvm/test/Transforms/GlobalOpt/atexit-dtor.ll

@@ -0,0 +1,54 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt < %s -S -passes=globalopt | FileCheck %s
+
+declare dso_local i32 @atexit(ptr)
+
+define dso_local void @empty_atexit_handler() {
+; CHECK-LABEL: define dso_local void @empty_atexit_handler() local_unnamed_addr {
+; CHECK-NEXT:    ret void
+;
+  ret void
+}
+
+; Check that `atexit` is removed if the handler is empty.
+; Check that a removed `atexit` call returns `0` which is the value that denotes success.
+define dso_local noundef i32 @register_atexit_handler() {
+; CHECK-LABEL: define dso_local noundef i32 @register_atexit_handler() local_unnamed_addr {
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @atexit(ptr @empty_atexit_handler)
+  ret i32 %1
+}
+
+declare dso_local void @declared_atexit_handler()
+
+; Check that an atexit handler with only a declaration is not removed.
+define dso_local noundef i32 @register_declared_atexit_handler() {
+; CHECK-LABEL: define dso_local noundef i32 @register_declared_atexit_handler() local_unnamed_addr {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @atexit(ptr @declared_atexit_handler)
+; CHECK-NEXT:    ret i32 [[TMP1]]
+;
+  %1 = call i32 @atexit(ptr @declared_atexit_handler)
+  ret i32 %1
+}
+
+declare dso_local void @external_exit_func()
+
+define dso_local void @nonempty_atexit_handler() {
+; CHECK-LABEL: define dso_local void @nonempty_atexit_handler() {
+; CHECK-NEXT:    call void @external_exit_func()
+; CHECK-NEXT:    ret void
+;
+  call void @external_exit_func()
+  ret void
+}
+
+; Check that an atexit handler that consists of any instructions other than `ret` is considered nonempty and not removed.
+define dso_local noundef i32 @register_nonempty_atexit_handler() {
+; CHECK-LABEL: define dso_local noundef i32 @register_nonempty_atexit_handler() local_unnamed_addr {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @atexit(ptr @nonempty_atexit_handler)
+; CHECK-NEXT:    ret i32 [[TMP1]]
+;
+  %1 = call i32 @atexit(ptr @nonempty_atexit_handler)
+  ret i32 %1
+}

llvm/test/tools/llvm-tli-checker/ps4-tli-check.yaml

@@ -34,21 +34,21 @@
 #
 # CHECK: << Total TLI yes SDK no:  8
 # CHECK: >> Total TLI no  SDK yes: 0
-# CHECK: == Total TLI yes SDK yes: 238
+# CHECK: == Total TLI yes SDK yes: 239
 #
 # WRONG_DETAIL: << TLI yes SDK no : '_ZdaPv' aka operator delete[](void*)
 # WRONG_DETAIL: >> TLI no  SDK yes: '_ZdaPvj' aka operator delete[](void*, unsigned int)
 # WRONG_DETAIL-COUNT-8: << TLI yes SDK no : {{.*}}__hot_cold_t
 # WRONG_SUMMARY: << Total TLI yes SDK no:  9{{$}}
 # WRONG_SUMMARY: >> Total TLI no  SDK yes: 1{{$}}
-# WRONG_SUMMARY: == Total TLI yes SDK yes: 237
+# WRONG_SUMMARY: == Total TLI yes SDK yes: 238
 #
 ## The -COUNT suffix doesn't care if there are too many matches, so check
 ## the exact count first; the two directives should add up to that.
 ## Yes, this means additions to TLI will fail this test, but the argument
 ## to -COUNT can't be an expression.
-# AVAIL: TLI knows 479 symbols, 246 available
-# AVAIL-COUNT-246: {{^}} available
+# AVAIL: TLI knows 480 symbols, 247 available
+# AVAIL-COUNT-247: {{^}} available
 # AVAIL-NOT:       {{^}} available
 # UNAVAIL-COUNT-233: not available
 # UNAVAIL-NOT:       not available
@@ -263,6 +263,10 @@ DynamicSymbols:
     Type:            STT_FUNC
     Section:         .text
     Binding:         STB_GLOBAL
+  - Name:            atexit
+    Type:            STT_FUNC
+    Section:         .text
+    Binding:         STB_GLOBAL
   - Name:            atof
     Type:            STT_FUNC
     Section:         .text

llvm/unittests/Analysis/TargetLibraryInfoTest.cpp

@@ -495,6 +495,8 @@ TEST_F(TargetLibraryInfoTest, ValidProto) {
       "declare i32 @__cxa_guard_acquire(%struct*)\n"
       "declare void @__cxa_guard_release(%struct*)\n"
 
+      "declare i32 @atexit(void ()*)\n"
+
       "declare i32 @__nvvm_reflect(i8*)\n"
 
       "declare i8* @__memcpy_chk(i8*, i8*, i64, i64)\n"