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

MaxEW707 · 2024-04-16T03:37:08Z

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.

llvmbot · 2024-04-16T03:37:41Z

@llvm/pr-subscribers-llvm-analysis

@llvm/pr-subscribers-llvm-transforms

Author: Max Winkler (MaxEW707)

Changes

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.

Full diff: https://github.com/llvm/llvm-project/pull/88836.diff

3 Files Affected:

(modified) llvm/include/llvm/Analysis/TargetLibraryInfo.def (+5)
(modified) llvm/lib/Transforms/IPO/GlobalOpt.cpp (+29-13)
(added) llvm/test/Transforms/GlobalOpt/atexit-dtor.ll (+46)

diff --git a/llvm/include/llvm/Analysis/TargetLibraryInfo.def b/llvm/include/llvm/Analysis/TargetLibraryInfo.def
index 37221eb9e47115..717693a7cf63c1 100644
--- a/llvm/include/llvm/Analysis/TargetLibraryInfo.def
+++ b/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)
diff --git a/llvm/lib/Transforms/IPO/GlobalOpt.cpp b/llvm/lib/Transforms/IPO/GlobalOpt.cpp
index da714c9a75701b..bf277efc892bff 100644
--- a/llvm/lib/Transforms/IPO/GlobalOpt.cpp
+++ b/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");
@@ -2321,14 +2322,14 @@ 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;
+  LibFunc F = Func;
   if (!TLI->has(F))
     return nullptr;
 
@@ -2340,17 +2341,27 @@ FindCXAAtExit(Module &M, function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
   TLI = &GetTLI(*Fn);
 
   // Make sure that the function has the correct prototype.
-  if (!TLI->getLibFunc(*Fn, F) || F != LibFunc_cxa_atexit)
+  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.
+static Function *
+FindCXAAtExit(Module &M, function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
+  return FindAtExitLibFunc(M, GetTLI, LibFunc_cxa_atexit);
+}
+
+static Function *
+FindAtExit(Module &M, function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
+  return FindAtExitLibFunc(M, GetTLI, LibFunc_atexit);
+}
+
+/// 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 cxxDtorIsEmpty(const Function &Fn) {
+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())
@@ -2366,7 +2377,7 @@ 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
@@ -2379,7 +2390,7 @@ 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
+  // This pass will look for calls to __cxa_atexit or atexit where the function is trivial
   // and remove them.
   bool Changed = false;
 
@@ -2393,14 +2404,17 @@ 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;
   }
@@ -2518,9 +2532,11 @@ 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 = FindCXAAtExit(M, GetTLI))
+      LocalChange |= OptimizeEmptyGlobalAtExitDtors(CXAAtExitFn, true);
+
+    if (Function *AtExitFn = FindAtExit(M, GetTLI))
+      LocalChange |= OptimizeEmptyGlobalAtExitDtors(AtExitFn, false);
 
     // Optimize IFuncs whose callee's are statically known.
     LocalChange |= OptimizeStaticIFuncs(M);
diff --git a/llvm/test/Transforms/GlobalOpt/atexit-dtor.ll b/llvm/test/Transforms/GlobalOpt/atexit-dtor.ll
new file mode 100644
index 00000000000000..f504fcc9ae29d3
--- /dev/null
+++ b/llvm/test/Transforms/GlobalOpt/atexit-dtor.ll
@@ -0,0 +1,46 @@
+; RUN: opt < %s -S -passes='cgscc(inline),function(early-cse),globalopt' | FileCheck %s
+
+%struct.A = type { i32 }
+
+$"??1A@@QEAA@XZ" = comdat any
+
+@"?g@@3UA@@A" = dso_local global %struct.A zeroinitializer, align 4
+@llvm.global_ctors = appending global [1 x { i32, ptr, ptr }] [{ i32, ptr, ptr } { i32 65535, ptr @_GLOBAL__sub_I_atexit-dtor, ptr null }]
+
+; CHECK-NOT: call i32 @atexit
+
+define internal void @"??__Eg@@YAXXZ"() {
+  %1 = call i32 @atexit(ptr @"??__Fg@@YAXXZ")
+  ret void
+}
+
+define linkonce_odr dso_local void @"??1A@@QEAA@XZ"(ptr noundef nonnull align 4 dereferenceable(4) %0) unnamed_addr #1 comdat align 2 {
+  ret void
+}
+
+define internal void @"??__Fg@@YAXXZ"() {
+  call void @"??1A@@QEAA@XZ"(ptr @"?g@@3UA@@A")
+  ret void
+}
+
+declare dso_local i32 @atexit(ptr)
+
+define internal void @_GLOBAL__sub_I_atexit-dtor() {
+  call void @"??__Eg@@YAXXZ"()
+  ret void
+}
+
+define dso_local void @atexit_handler() {
+  ret void
+}
+
+; CHECK-NOT: call i32 @atexit
+
+; Check that a removed `atexit` call returns `0` which is the value that denotes success.
+define dso_local noundef i32 @register_atexit_handler() {
+  %1 = alloca i32, align 4
+  store i32 0, ptr %1, align 4
+  %2 = call i32 @atexit(ptr @"atexit_handler")
+; CHECK: ret i32 0
+  ret i32 %2
+}

github-actions · 2024-04-16T03:40:10Z

✅ With the latest revision this PR passed the C/C++ code formatter.

MaxEW707 · 2024-04-17T01:26:06Z

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

@@ -263,6 +263,10 @@ DynamicSymbols:
    Type:            STT_FUNC
    Section:         .text
    Binding:         STB_GLOBAL
+  - Name:            atexit


@pogo59 @sks75220 Let me know for ps4/ps5 if you prefer that this library function be set unavailable.

atexit should be available for PS4/PS5, thanks!

llvm/lib/Transforms/IPO/GlobalOpt.cpp

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

nikic

LGTM

llvmbot added llvm:analysis llvm:transforms labels Apr 16, 2024

MaxEW707 added 8 commits April 16, 2024 20:15

Remove empty atexit destructors

cb1616a

formatting

d5cd155

eol

f5e276d

unit test for explicit at exit registration

d38ec0c

eol

dfd476b

formatting

a3a8419

fix unit test

8bedd9c

Add atexit to ps tli checker

f1b585f

MaxEW707 force-pushed the mew/atexit-dtor-opt branch from c2b100f to f1b585f Compare April 17, 2024 01:14

fix formatting

08f30d5

MaxEW707 commented Apr 17, 2024

View reviewed changes

MaxEW707 requested a review from nikic April 20, 2024 20:13

nikic reviewed Apr 25, 2024

View reviewed changes

MaxEW707 added 3 commits April 24, 2024 23:42

Remove static functions; Feedback variable scoping

e9b1eb5

autogenerate check statements for unit test

00c6aed

fix formatting

23ff313

nikic reviewed Apr 26, 2024

View reviewed changes

llvm/test/Transforms/GlobalOpt/atexit-dtor.ll Outdated Show resolved Hide resolved

llvm/test/Transforms/GlobalOpt/atexit-dtor.ll Outdated Show resolved Hide resolved

llvm/test/Transforms/GlobalOpt/atexit-dtor.ll Outdated Show resolved Hide resolved

add negative unit tests

cedb55f

nikic approved these changes Apr 27, 2024

View reviewed changes

Rebuild due to windows agent out of heap space error from msvc

8abf12a

MaxEW707 merged commit ce3485a into llvm:main Apr 30, 2024
4 checks passed

MaxEW707 deleted the mew/atexit-dtor-opt branch April 30, 2024 00:29

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

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

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

Uh oh!

MaxEW707 commented Apr 16, 2024

Uh oh!

llvmbot commented Apr 16, 2024 •

edited

Loading

Uh oh!

github-actions bot commented Apr 16, 2024 •

edited

Loading

Uh oh!

MaxEW707 Apr 17, 2024 •

edited

Loading

Uh oh!

pogo59 Apr 17, 2024

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

nikic left a comment

Uh oh!

Uh oh!

Uh oh!

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

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

Uh oh!

Conversation

MaxEW707 commented Apr 16, 2024

Uh oh!

llvmbot commented Apr 16, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Apr 16, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

MaxEW707 Apr 17, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pogo59 Apr 17, 2024

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

nikic left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

llvmbot commented Apr 16, 2024 •

edited

Loading

github-actions bot commented Apr 16, 2024 •

edited

Loading

MaxEW707 Apr 17, 2024 •

edited

Loading