[Clang] Mark this pointer in destructors dead_on_return#166276
Merged
boomanaiden154 merged 13 commits intollvm:mainfrom Feb 6, 2026
Merged
[Clang] Mark this pointer in destructors dead_on_return#166276boomanaiden154 merged 13 commits intollvm:mainfrom
boomanaiden154 merged 13 commits intollvm:mainfrom
Conversation
This helps to clean up any dead stores that come up at the end of the destructor. The motivating example was a refactoring in libc++'s basic_string implementation in 8dae17b that added a zeroing store into the destructor, causing a large performance regression on an internal workload.
boomanaiden154
force-pushed
the
clang-destructor-dead-on-return
branch
from
7fb6339 to
7a3dec4
Compare
llvmbot
added
clang
Member
|
@llvm/pr-subscribers-clang-driver @llvm/pr-subscribers-clang Author: Aiden Grossman (boomanaiden154) ChangesThis helps to clean up any dead stores that come up at the end of the destructor. The motivating example was a refactoring in libc++'s basic_string implementation in 8dae17b that added a zeroing store into the destructor, causing a large performance regression on an internal workload. We also saw a ~0.2% performance increase on an internal server workload when enabling this. I also tested this against all of the non-flaky tests in our large C++ codebase and found a minimal number of issues that all happened to be in user code. Patch is 5.29 MiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/166276.diff 109 Files Affected:
diff --git a/clang/lib/CodeGen/CGCall.cpp b/clang/lib/CodeGen/CGCall.cpp
index efacb3cc04c01..ee6e13fd1c1a5 100644
--- a/clang/lib/CodeGen/CGCall.cpp
+++ b/clang/lib/CodeGen/CGCall.cpp
@@ -2767,7 +2767,8 @@ void CodeGenModule::ConstructAttributeList(StringRef Name,
}
// Apply `nonnull`, `dereferenceable(N)` and `align N` to the `this` argument,
- // unless this is a thunk function.
+ // unless this is a thunk function. Add dead_on_return to the `this` argument
+ // in base class destructors to aid in DSE.
// FIXME: fix this properly, https://reviews.llvm.org/D100388
if (FI.isInstanceMethod() && !IRFunctionArgs.hasInallocaArg() &&
!FI.arg_begin()->type->isVoidPointerType() && !IsThunk) {
@@ -2800,6 +2801,15 @@ void CodeGenModule::ConstructAttributeList(StringRef Name,
.getAsAlign();
Attrs.addAlignmentAttr(Alignment);
+ if (isa_and_nonnull<CXXDestructorDecl>(
+ CalleeInfo.getCalleeDecl().getDecl())) {
+ auto *ClassDecl = dyn_cast<CXXRecordDecl>(
+ CalleeInfo.getCalleeDecl().getDecl()->getDeclContext());
+ if (ClassDecl->getNumBases() == 0 && ClassDecl->getNumVBases() == 0) {
+ Attrs.addAttribute(llvm::Attribute::DeadOnReturn);
+ }
+ }
+
ArgAttrs[IRArgs.first] = llvm::AttributeSet::get(getLLVMContext(), Attrs);
}
diff --git a/clang/test/CodeGen/paren-list-agg-init.cpp b/clang/test/CodeGen/paren-list-agg-init.cpp
index e30777ecc07d6..561bf2b5eb9c4 100644
--- a/clang/test/CodeGen/paren-list-agg-init.cpp
+++ b/clang/test/CodeGen/paren-list-agg-init.cpp
@@ -394,9 +394,9 @@ namespace gh61145 {
// a.k.a. Vec::Vec(Vec&&)
// CHECK-NEXT: call void @_ZN7gh611453VecC1EOS0_(ptr noundef nonnull align 1 dereferenceable(1) [[AGG_TMP_ENSURED]], ptr noundef nonnull align 1 dereferenceable(1) [[V]])
// a.k.a. S1::~S1()
- // CHECK-NEXT: call void @_ZN7gh611452S1D1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[AGG_TMP_ENSURED]])
+ // CHECK-NEXT: call void @_ZN7gh611452S1D1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(1) [[AGG_TMP_ENSURED]])
// a.k.a.Vec::~Vec()
- // CHECK-NEXT: call void @_ZN7gh611453VecD1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[V]])
+ // CHECK-NEXT: call void @_ZN7gh611453VecD1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(1) [[V]])
// CHECK-NEXT: ret void
template <int I>
void make1() {
@@ -416,9 +416,9 @@ namespace gh61145 {
// CHECK-NEXT: [[C:%.*c.*]] = getelementptr inbounds nuw [[STRUCT_S2]], ptr [[AGG_TMP_ENSURED]], i32 0, i32
// CHECK-NEXT: store i8 0, ptr [[C]], align 1
// a.k.a. S2::~S2()
- // CHECK-NEXT: call void @_ZN7gh611452S2D1Ev(ptr noundef nonnull align 1 dereferenceable(2) [[AGG_TMP_ENSURED]])
+ // CHECK-NEXT: call void @_ZN7gh611452S2D1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(2) [[AGG_TMP_ENSURED]])
// a.k.a. Vec::~Vec()
- // CHECK-NEXT: call void @_ZN7gh611453VecD1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[V]])
+ // CHECK-NEXT: call void @_ZN7gh611453VecD1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(1) [[V]])
// CHECK-NEXT: ret void
template <int I>
void make2() {
diff --git a/clang/test/CodeGen/temporary-lifetime.cpp b/clang/test/CodeGen/temporary-lifetime.cpp
index 04087292b2c70..44d1235f15c86 100644
--- a/clang/test/CodeGen/temporary-lifetime.cpp
+++ b/clang/test/CodeGen/temporary-lifetime.cpp
@@ -24,12 +24,12 @@ void Test1() {
// CHECK-DTOR: call void @llvm.lifetime.start.p0(ptr nonnull %[[ADDR:.+]])
// CHECK-DTOR: call void @_ZN1AC1Ev(ptr nonnull {{[^,]*}} %[[VAR:[^ ]+]])
// CHECK-DTOR: call void @_Z3FooIRK1AEvOT_
- // CHECK-DTOR: call void @_ZN1AD1Ev(ptr nonnull {{[^,]*}} %[[VAR]])
+ // CHECK-DTOR: call void @_ZN1AD1Ev(ptr dead_on_return nonnull {{[^,]*}} %[[VAR]])
// CHECK-DTOR: call void @llvm.lifetime.end.p0(ptr nonnull %[[ADDR]])
// CHECK-DTOR: call void @llvm.lifetime.start.p0(ptr nonnull %[[ADDR:.+]])
// CHECK-DTOR: call void @_ZN1AC1Ev(ptr nonnull {{[^,]*}} %[[VAR:[^ ]+]])
// CHECK-DTOR: call void @_Z3FooIRK1AEvOT_
- // CHECK-DTOR: call void @_ZN1AD1Ev(ptr nonnull {{[^,]*}} %[[VAR]])
+ // CHECK-DTOR: call void @_ZN1AD1Ev(ptr dead_on_return nonnull {{[^,]*}} %[[VAR]])
// CHECK-DTOR: call void @llvm.lifetime.end.p0(ptr nonnull %[[ADDR]])
// CHECK-DTOR: }
@@ -61,9 +61,9 @@ void Test2() {
// CHECK-DTOR: call void @llvm.lifetime.start.p0(ptr nonnull %[[ADDR2:.+]])
// CHECK-DTOR: call void @_ZN1AC1Ev(ptr nonnull {{[^,]*}} %[[VAR2:[^ ]+]])
// CHECK-DTOR: call void @_Z3FooIRK1AEvOT_
- // CHECK-DTOR: call void @_ZN1AD1Ev(ptr nonnull {{[^,]*}} %[[VAR2]])
+ // CHECK-DTOR: call void @_ZN1AD1Ev(ptr dead_on_return nonnull {{[^,]*}} %[[VAR2]])
// CHECK-DTOR: call void @llvm.lifetime.end.p0(ptr nonnull %[[ADDR2]])
- // CHECK-DTOR: call void @_ZN1AD1Ev(ptr nonnull {{[^,]*}} %[[VAR1]])
+ // CHECK-DTOR: call void @_ZN1AD1Ev(ptr dead_on_return nonnull {{[^,]*}} %[[VAR1]])
// CHECK-DTOR: call void @llvm.lifetime.end.p0(ptr nonnull %[[ADDR1]])
// CHECK-DTOR: }
@@ -155,12 +155,12 @@ void Test7() {
// CHECK-DTOR: call void @llvm.lifetime.start.p0(ptr nonnull %[[ADDR:.+]])
// CHECK-DTOR: call void @_Z3BazI1AET_v({{.*}} %[[SLOT:[^ ]+]])
// CHECK-DTOR: call void @_Z3FooI1AEvOT_({{.*}} %[[SLOT]])
- // CHECK-DTOR: call void @_ZN1AD1Ev(ptr nonnull {{[^,]*}} %[[SLOT]])
+ // CHECK-DTOR: call void @_ZN1AD1Ev(ptr dead_on_return nonnull {{[^,]*}} %[[SLOT]])
// CHECK-DTOR: call void @llvm.lifetime.end.p0(ptr nonnull %[[ADDR]])
// CHECK-DTOR: call void @llvm.lifetime.start.p0(ptr nonnull %[[ADDR:.+]])
// CHECK-DTOR: call void @_Z3BazI1AET_v({{.*}} %[[SLOT:[^ ]+]])
// CHECK-DTOR: call void @_Z3FooI1AEvOT_({{.*}} %[[SLOT]])
- // CHECK-DTOR: call void @_ZN1AD1Ev(ptr nonnull {{[^,]*}} %[[SLOT]])
+ // CHECK-DTOR: call void @_ZN1AD1Ev(ptr dead_on_return nonnull {{[^,]*}} %[[SLOT]])
// CHECK-DTOR: call void @llvm.lifetime.end.p0(ptr nonnull %[[ADDR]])
// CHECK-DTOR: }
Foo(Baz<A>());
diff --git a/clang/test/CodeGenCXX/amdgcn-automatic-variable.cpp b/clang/test/CodeGenCXX/amdgcn-automatic-variable.cpp
index 3c2a624bd4f95..e05f8133321c7 100644
--- a/clang/test/CodeGenCXX/amdgcn-automatic-variable.cpp
+++ b/clang/test/CodeGenCXX/amdgcn-automatic-variable.cpp
@@ -75,7 +75,7 @@ int x;
// CHECK-NEXT: [[A:%.*]] = alloca [[CLASS_A:%.*]], align 4, addrspace(5)
// CHECK-NEXT: [[A_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[A]] to ptr
// CHECK-NEXT: call void @_ZN1AC1Ev(ptr noundef nonnull align 4 dereferenceable(4) [[A_ASCAST]])
-// CHECK-NEXT: call void @_ZN1AD1Ev(ptr noundef nonnull align 4 dereferenceable(4) [[A_ASCAST]])
+// CHECK-NEXT: call void @_ZN1AD1Ev(ptr dead_on_return noundef nonnull align 4 dereferenceable(4) [[A_ASCAST]])
// CHECK-NEXT: ret void
//
void func3() {
diff --git a/clang/test/CodeGenCXX/amdgcn-func-arg.cpp b/clang/test/CodeGenCXX/amdgcn-func-arg.cpp
index a5f83dc91b038..bc20c33ec4d0f 100644
--- a/clang/test/CodeGenCXX/amdgcn-func-arg.cpp
+++ b/clang/test/CodeGenCXX/amdgcn-func-arg.cpp
@@ -43,9 +43,9 @@ void func_with_indirect_arg(A a) {
// CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 4 [[AGG_TMP_ASCAST]], ptr align 4 [[A_ASCAST]], i64 4, i1 false)
// CHECK-NEXT: [[AGG_TMP_ASCAST_ASCAST:%.*]] = addrspacecast ptr [[AGG_TMP_ASCAST]] to ptr addrspace(5)
// CHECK-NEXT: call void @_Z22func_with_indirect_arg1A(ptr addrspace(5) noundef [[AGG_TMP_ASCAST_ASCAST]])
-// CHECK-NEXT: call void @_ZN1AD1Ev(ptr noundef nonnull align 4 dereferenceable(4) [[AGG_TMP_ASCAST]])
+// CHECK-NEXT: call void @_ZN1AD1Ev(ptr dead_on_return noundef nonnull align 4 dereferenceable(4) [[AGG_TMP_ASCAST]])
// CHECK-NEXT: call void @_Z17func_with_ref_argR1A(ptr noundef nonnull align 4 dereferenceable(4) [[A_ASCAST]])
-// CHECK-NEXT: call void @_ZN1AD1Ev(ptr noundef nonnull align 4 dereferenceable(4) [[A_ASCAST]])
+// CHECK-NEXT: call void @_ZN1AD1Ev(ptr dead_on_return noundef nonnull align 4 dereferenceable(4) [[A_ASCAST]])
// CHECK-NEXT: ret void
//
void test_indirect_arg_auto() {
@@ -61,7 +61,7 @@ void test_indirect_arg_auto() {
// CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 4 [[AGG_TMP_ASCAST]], ptr align 4 addrspacecast (ptr addrspace(1) @g_a to ptr), i64 4, i1 false)
// CHECK-NEXT: [[AGG_TMP_ASCAST_ASCAST:%.*]] = addrspacecast ptr [[AGG_TMP_ASCAST]] to ptr addrspace(5)
// CHECK-NEXT: call void @_Z22func_with_indirect_arg1A(ptr addrspace(5) noundef [[AGG_TMP_ASCAST_ASCAST]])
-// CHECK-NEXT: call void @_ZN1AD1Ev(ptr noundef nonnull align 4 dereferenceable(4) [[AGG_TMP_ASCAST]])
+// CHECK-NEXT: call void @_ZN1AD1Ev(ptr dead_on_return noundef nonnull align 4 dereferenceable(4) [[AGG_TMP_ASCAST]])
// CHECK-NEXT: call void @_Z17func_with_ref_argR1A(ptr noundef nonnull align 4 dereferenceable(4) addrspacecast (ptr addrspace(1) @g_a to ptr))
// CHECK-NEXT: ret void
//
diff --git a/clang/test/CodeGenCXX/control-flow-in-stmt-expr.cpp b/clang/test/CodeGenCXX/control-flow-in-stmt-expr.cpp
index 4eafa720e0cb4..a764ba31539eb 100644
--- a/clang/test/CodeGenCXX/control-flow-in-stmt-expr.cpp
+++ b/clang/test/CodeGenCXX/control-flow-in-stmt-expr.cpp
@@ -217,7 +217,7 @@ void ArrayInit() {
// CHECK: [[ARRAY_DESTROY_BODY2]]:
// CHECK-NEXT: %arraydestroy.elementPast = phi ptr [ %1, %cleanup ], [ %arraydestroy.element, %[[ARRAY_DESTROY_BODY2]] ]
// CHECK-NEXT: %arraydestroy.element = getelementptr inbounds %struct.Printy, ptr %arraydestroy.elementPast, i64 -1
- // CHECK-NEXT: call void @_ZN6PrintyD1Ev(ptr noundef nonnull align 8 dereferenceable(8) %arraydestroy.element)
+ // CHECK-NEXT: call void @_ZN6PrintyD1Ev(ptr dead_on_return noundef nonnull align 8 dereferenceable(8) %arraydestroy.element)
// CHECK-NEXT: %arraydestroy.done = icmp eq ptr %arraydestroy.element, %arr
// CHECK-NEXT: br i1 %arraydestroy.done, label %[[ARRAY_DESTROY_DONE2]], label %[[ARRAY_DESTROY_BODY2]]
@@ -265,7 +265,7 @@ void ArraySubobjects() {
// CHECK: [[ARRAY_DESTROY_BODY]]:
// CHECK-NEXT: %arraydestroy.elementPast = phi ptr [ %0, %if.then ], [ %arraydestroy.element, %[[ARRAY_DESTROY_BODY]] ]
// CHECK-NEXT: %arraydestroy.element = getelementptr inbounds %struct.Printy, ptr %arraydestroy.elementPast, i64 -1
- // CHECK-NEXT: call void @_ZN6PrintyD1Ev(ptr noundef nonnull align 8 dereferenceable(8) %arraydestroy.element)
+ // CHECK-NEXT: call void @_ZN6PrintyD1Ev(ptr dead_on_return noundef nonnull align 8 dereferenceable(8) %arraydestroy.element)
// CHECK-NEXT: %arraydestroy.done = icmp eq ptr %arraydestroy.element, %arr2
// CHECK-NEXT: br i1 %arraydestroy.done, label %[[ARRAY_DESTROY_DONE]], label %[[ARRAY_DESTROY_BODY]]
@@ -277,7 +277,7 @@ void ArraySubobjects() {
// CHECK: [[ARRAY_DESTROY_BODY2]]:
// CHECK-NEXT: %arraydestroy.elementPast4 = phi ptr [ %1, %[[ARRAY_DESTROY_DONE]] ], [ %arraydestroy.element5, %[[ARRAY_DESTROY_BODY2]] ]
// CHECK-NEXT: %arraydestroy.element5 = getelementptr inbounds %struct.Printy, ptr %arraydestroy.elementPast4, i64 -1
- // CHECK-NEXT: call void @_ZN6PrintyD1Ev(ptr noundef nonnull align 8 dereferenceable(8) %arraydestroy.element5)
+ // CHECK-NEXT: call void @_ZN6PrintyD1Ev(ptr dead_on_return noundef nonnull align 8 dereferenceable(8) %arraydestroy.element5)
// CHECK-NEXT: %arraydestroy.done6 = icmp eq ptr %arraydestroy.element5, [[ARRAY_BEGIN]]
// CHECK-NEXT: br i1 %arraydestroy.done6, label %[[ARRAY_DESTROY_DONE2:.+]], label %[[ARRAY_DESTROY_BODY2]]
@@ -384,7 +384,7 @@ void NewArrayInit() {
// CHECK: arraydestroy.body:
// CHECK-NEXT: %arraydestroy.elementPast = phi ptr [ %{{.*}}, %if.then ], [ %arraydestroy.element, %arraydestroy.body ]
// CHECK-NEXT: %arraydestroy.element = getelementptr inbounds %struct.Printy, ptr %arraydestroy.elementPast, i64 -1
- // CHECK-NEXT: call void @_ZN6PrintyD1Ev(ptr noundef nonnull align 8 dereferenceable(8) %arraydestroy.element)
+ // CHECK-NEXT: call void @_ZN6PrintyD1Ev(ptr dead_on_return noundef nonnull align 8 dereferenceable(8) %arraydestroy.element)
// CHECK-NEXT: %arraydestroy.done = icmp eq ptr %arraydestroy.element, %0
// CHECK-NEXT: br i1 %arraydestroy.done, label %arraydestroy.done{{.*}}, label %arraydestroy.body
diff --git a/clang/test/CodeGenCXX/cxx2a-destroying-delete.cpp b/clang/test/CodeGenCXX/cxx2a-destroying-delete.cpp
index 24b1a4dd42977..af29120feb5bb 100644
--- a/clang/test/CodeGenCXX/cxx2a-destroying-delete.cpp
+++ b/clang/test/CodeGenCXX/cxx2a-destroying-delete.cpp
@@ -41,11 +41,11 @@ void glob_delete_A(A *a) { ::delete a; }
// CHECK: icmp eq ptr %[[a]], null
// CHECK: br i1
-// CHECK-ITANIUM: call void @_ZN1AD1Ev(ptr noundef nonnull align 8 dereferenceable(8) %[[a]])
+// CHECK-ITANIUM: call void @_ZN1AD1Ev(ptr dead_on_return noundef nonnull align 8 dereferenceable(8) %[[a]])
// CHECK-ITANIUM-NEXT: call void @_ZdlPvm(ptr noundef %[[a]], i64 noundef 8)
-// CHECK-MSABI64: call void @"??1A@@QEAA@XZ"(ptr noundef nonnull align 8 dereferenceable(8) %[[a]])
+// CHECK-MSABI64: call void @"??1A@@QEAA@XZ"(ptr dead_on_return noundef nonnull align 8 dereferenceable(8) %[[a]])
// CHECK-MSABI64-NEXT: call void @"??3@YAXPEAX_K@Z"(ptr noundef %[[a]], i64 noundef 8)
-// CHECK-MSABI32: call x86_thiscallcc void @"??1A@@QAE@XZ"(ptr noundef nonnull align 4 dereferenceable(4) %[[a]])
+// CHECK-MSABI32: call x86_thiscallcc void @"??1A@@QAE@XZ"(ptr dead_on_return noundef nonnull align 4 dereferenceable(4) %[[a]])
// CHECK-MSABI32-NEXT: call void @"??3@YAXPAXI@Z"(ptr noundef %[[a]], i32 noundef 4)
struct B {
diff --git a/clang/test/CodeGenCXX/for-range.cpp b/clang/test/CodeGenCXX/for-range.cpp
index 088a34647c374..b9706855f658c 100644
--- a/clang/test/CodeGenCXX/for-range.cpp
+++ b/clang/test/CodeGenCXX/for-range.cpp
@@ -53,7 +53,7 @@ extern B array[5];
// CHECK: for.body:
// CHECK-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__BEGIN1]], align 8
// CHECK-NEXT: call void @_ZN1BC1ERKS_(ptr noundef nonnull align 1 dereferenceable(1) [[B]], ptr noundef nonnull align 1 dereferenceable(1) [[TMP2]])
-// CHECK-NEXT: call void @_ZN1BD1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[B]]) #[[ATTR3:[0-9]+]]
+// CHECK-NEXT: call void @_ZN1BD1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(1) [[B]]) #[[ATTR3:[0-9]+]]
// CHECK-NEXT: br label [[FOR_INC:%.*]]
// CHECK: for.inc:
// CHECK-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__BEGIN1]], align 8
@@ -61,7 +61,7 @@ extern B array[5];
// CHECK-NEXT: store ptr [[INCDEC_PTR]], ptr [[__BEGIN1]], align 8
// CHECK-NEXT: br label [[FOR_COND]]
// CHECK: for.end:
-// CHECK-NEXT: call void @_ZN1AD1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[A]]) #[[ATTR3]]
+// CHECK-NEXT: call void @_ZN1AD1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(1) [[A]]) #[[ATTR3]]
// CHECK-NEXT: ret void
//
void for_array() {
@@ -81,10 +81,10 @@ void for_array() {
// CHECK-NEXT: call void @_ZN1AC1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[A]])
// CHECK-NEXT: call void @_ZN1CC1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[REF_TMP]])
// CHECK-NEXT: store ptr [[REF_TMP]], ptr [[__RANGE1]], align 8
-// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE1]], align 8
+// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE1]], align 8, !nonnull [[META2:![0-9]+]]
// CHECK-NEXT: [[CALL:%.*]] = call noundef ptr @_Z5beginR1C(ptr noundef nonnull align 1 dereferenceable(1) [[TMP0]])
// CHECK-NEXT: store ptr [[CALL]], ptr [[__BEGIN1]], align 8
-// CHECK-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE1]], align 8
+// CHECK-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE1]], align 8, !nonnull [[META2]]
// CHECK-NEXT: [[CALL1:%.*]] = call noundef ptr @_Z3endR1C(ptr noundef nonnull align 1 dereferenceable(1) [[TMP1]])
// CHECK-NEXT: store ptr [[CALL1]], ptr [[__END1]], align 8
// CHECK-NEXT: br label [[FOR_COND:%.*]]
@@ -94,12 +94,12 @@ void for_array() {
// CHECK-NEXT: [[CMP:%.*]] = icmp ne ptr [[TMP2]], [[TMP3]]
// CHECK-NEXT: br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_COND_CLEANUP:%.*]]
// CHECK: for.cond.cleanup:
-// CHECK-NEXT: call void @_ZN1CD1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[REF_TMP]]) #[[ATTR3]]
+// CHECK-NEXT: call void @_ZN1CD1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(1) [[REF_TMP]]) #[[ATTR3]]
// CHECK-NEXT: br label [[FOR_END:%.*]]
// CHECK: for.body:
// CHECK-NEXT: [[TMP4:%.*]] = load ptr, ptr [[__BEGIN1]], align 8
// CHECK-NEXT: call void @_ZN1BC1ERKS_(ptr noundef nonnull align 1 dereferenceable(1) [[B]], ptr noundef nonnull align 1 dereferenceable(1) [[TMP4]])
-// CHECK-NEXT: call void @_ZN1BD1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[B]]) #[[ATTR3]]
+// CHECK-NEXT: call void @_ZN1BD1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(1) [[B]]) #[[ATTR3]]
// CHECK-NEXT: br label [[FOR_INC:%.*]]
// CHECK: for.inc:
// CHECK-NEXT: [[TMP5:%.*]] = load ptr, ptr [[__BEGIN1]], align 8
@@ -107,7 +107,7 @@ void for_array() {
// CHECK-NEXT: store ptr [[INCDEC_PTR]], ptr [[__BEGIN1]], align 8
// CHECK-NEXT: br label [[FOR_COND]]
// CHECK: for.end:
-// CHECK-NEXT: call void @_ZN1AD1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[A]]) #[[ATTR3]]
+// CHECK-NEXT: call void @_ZN1AD1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(1) [[A]]) #[[ATTR3]]
// CHECK-NEXT: ret void
//
void for_range() {
@@ -127,10 +127,10 @@ void for_range() {
// CHECK-NEXT: call void @_ZN1AC1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[A]])
// CHECK-NEXT: call void @_ZN1DC1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[REF_TMP]])
// CHECK-NEXT: store ptr [[REF_TMP]], ptr [[__RANGE1]], align 8
-// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE1]], align 8
+// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE1]], align 8, !nonnull [[META2]]
// CHECK-NEXT: [[CALL:%.*]] = call noundef ptr @_ZN1D5beginEv(ptr noundef nonnull align 1 dereferenceable(1) [[TMP0]])
// CHECK-NEXT: store ptr [[CALL]], ptr [[__BEGIN1]], align 8
-// CHECK-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE1]], align 8
+// CHECK-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE1]], align 8, !nonnull [[META2]]
// CHECK-NEXT: [[CALL1:%.*]] = call noundef ptr @_ZN1D3endEv(ptr noundef nonnull align 1 dereferenceable(1) [[TMP1]])
// CHECK-NEXT: store ptr [[CALL1]], ptr [[__END1]], align 8
// CHECK-NEXT: br label [[FOR_COND:%.*]]
@@ -140,12 +140,12 @@ void for_range() {
// CHECK-NEXT: [[CMP:%.*]] = icmp ne ptr [[TMP2]], [[TMP3]]
// CHECK-NEXT: br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_COND_CLEANUP:%.*]]
// CHECK: for.cond.cleanup:
-// CHECK-NEXT: call void @_ZN1DD1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[REF_TMP]]) #[[ATTR3]]
+// CHECK-NEXT: call void @_ZN1DD1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(1) [[REF_TMP]]) #[[ATTR3]]
// CHECK-NEXT: br label [[FOR_END:%.*]]
// CHECK: for.body:
// CHECK-NEXT: [[TMP4:%.*]] = load ptr, ptr [[__BEGIN1]], align 8
// CHECK-NEXT: call void @_ZN1BC1ERKS_(ptr noundef nonnull align 1 dereferenceable(1) [[B]], ptr noundef nonnull align 1 dereferenceable(1) [[TMP4]])
-// CHECK-NEXT: call void @_ZN1BD1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[B]]) #[[ATTR3]]
+// CHECK-NEXT: call void @_ZN1BD1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(1) [[B]]) #[[ATTR3]]
// CHECK-NEXT: br label [[FOR_INC:%.*]]
// CHECK: for.inc:
// CHECK-NEXT: [[TMP5:%.*]] = load ptr, ptr [[__BEGIN1]], align 8
@@ -153,7 +153,7 @@ void for_range() {
// CHECK-NEXT: store ptr [[INCDEC_PTR]], ptr [[__BEGIN1]], align 8
// CHECK-NEXT: br label [[FOR_COND]]
// CHECK: for.end:
-// CHECK-NEXT: call void @_ZN1AD1Ev(ptr noundef nonnull align 1 dereferenceable(1) [[A]]) #[[ATTR3]]
+// CHECK-NEXT: call void @_ZN1AD1Ev(ptr dead_on_return noundef nonnull align 1 dereferenceable(1) [[A]]) #[[ATTR3]]
// CHECK-NEXT: ret void
//
void for_member_range() {
diff --git a/clang/test/CodeGenCXX/gh62818.cpp b/clang/test/CodeGenCXX/gh62818.cpp
index ec91b40fca077..f903679cd6b68 100644
--- a/clang/test/CodeGenCXX/gh62818.cpp
+++ b/clang/test/CodeGenCXX/gh62818.c...
[truncated]
|
Contributor
Author
|
The actual change is in the first commit (7a3dec4). I've separated the test changes out into 6ec0952 to hopefully make review a bit easier. CC @philnik777 Who came up with the idea and requested this. |
boomanaiden154
requested review from
AaronBallman,
antoniofrighetto,
efriedma-quic,
jhuber6 and
rnk
Collaborator
|
This optimization exploits the fact that it's undefined behavior to read from an object after its been destroyed. Given the overall shift in how the industry feels about compilers exploiting undefined behavior, I want to push to add an flag to control this. Think of the people who use I'd also like to better understand why base classes matter for this annotation. Until very recently, |
rnk
reviewed
Dec 1, 2025
Collaborator
rnk
left a comment
rnk
left a comment
There was a problem hiding this comment.
Overall, we should do it, this is a valuable optimization. (Commenting twice to push out the inline comments).
clang/lib/CodeGen/CGCall.cpp
Outdated
| .getAsAlign(); | ||
| Attrs.addAlignmentAttr(Alignment); | ||
|
|
||
| if (isa_and_nonnull<CXXDestructorDecl>( |
Collaborator
There was a problem hiding this comment.
Code golf: You can CSE the CalleeInfo.getCalleeDecl().getDecl() if you use dyn_cast_or_nonnull.
Contributor
Author
There was a problem hiding this comment.
Done. Ended up using dyn_cast_if_present.
clang/lib/CodeGen/CGCall.cpp
Outdated
| CalleeInfo.getCalleeDecl().getDecl())) { | ||
| auto *ClassDecl = dyn_cast<CXXRecordDecl>( | ||
| CalleeInfo.getCalleeDecl().getDecl()->getDeclContext()); | ||
| if (ClassDecl->getNumBases() == 0 && ClassDecl->getNumVBases() == 0) { |
Collaborator
There was a problem hiding this comment.
Why do we have to limit this to only destructors of classes with no bases? Whatever the reason (caution, incremental change, etc), comments here would be appreciated.
Contributor
There was a problem hiding this comment.
Virtual base subobjects aren't dead after a base subobject destructor call, but yeah, I can't think of a reason to limit this because of non-virtual bases alone. And even virtual bases are dead after a complete-object destructor call.
Contributor
Author
There was a problem hiding this comment.
For now, just to be conservative while we get implementation experience. I've added a TODO for myself to remove this after we do more testing.
Collaborator
efriedma-quic
left a comment
efriedma-quic
left a comment
There was a problem hiding this comment.
Is the potential undefined behavior here checked by msan? Do we need to disable adding this attribute if msan is enabled?
Contributor
Author
Yes, use-after-destroy is checked by msan, but I don't believe we need to explicitly disable anything here to handle msan. Based on my understand, the msan instrumentation will add a call to CC @thurstond Who is more familiar with the internals of msan and would be able to confirm. |
Contributor
Author
I was mostly concerned about cases like https://godbolt.org/z/fq4fWfqKv, thinking that we might eliminate stores that would otherwise propagate into the base destructors. I didn't realize that the invocation of the more-base destructor just happens as a call at the end of the derived class destructor, so is a non-issue. Besides that, I wouldn't mind being conservative initially and making sure that everything ships internally as smoothly as the initial testing that I've done before making this optimization more broad. It doesn't seem like there would be large fallout, but it doesn't seem like there's a downside to doing things incrementally. |
Contributor
|
You're correct to be conservative about virtual bases, because they will generally still be live out of the base subobject destructor. I think it's worth doing this for non-virtual bases just to avoid needing an excessive number of controlling flags, since otherwise I assume people might want a staging flag that disables this just for classes with bases. |
llvmbot
added
clang:driver
Contributor
|
GCC has a related |
Contributor
This is my understanding as well.
Yep, shadow memory is opaque to the clang backend - the shadow is just a big blob of memory (~16TB on x86-64), not a bunch of objects. |
Contributor
Author
Thanks for confirming. I've constructed https://godbolt.org/z/nKPfhPxsx which I believe demonstrates the problem there (we end up destroying Given you said this is in general, it sounds like there are cases where virtual bases are not live out of the base subobject destructor? |
Contributor
Author
Reusing the existing flag name sounds reasonable enough to me, but in #40040 it seems like people had opinions on the flag naming, with @zygoloid suggesting Regarding #24952, I guess I did not end up hitting that one because of the intentional conservativeness around base classes (I also didn't do a bootstrapping build with LTO). |
Contributor
Author
I've left this as a TODO, mainly so I can make sure that our internal release process doesn't catch many more issues than how I tested this. Then I plan on doing another round of testing without the check on non-virtual bases and then removing the TODO. We don't need more selective flags than this and I don't see why other users couldn't just be pointed at an opt-out flag. Is this strategy okay or would you like to see the non-virtual base check removed before landing? |
🪟 Windows x64 Test Results
✅ The build succeeded and all tests passed. |
🐧 Linux x64 Test Results
✅ The build succeeded and all tests passed. |
Contributor
Author
|
Bump on this when reviewers get a chance. Thanks! |
zygoloid
reviewed
Feb 2, 2026
zygoloid
reviewed
Feb 2, 2026
Collaborator
zygoloid
left a comment
zygoloid
left a comment
There was a problem hiding this comment.
LGTM; I believe this is correct from a language rules and ABI perspective (but you should get approval from someone who's more involved with Clang these days).
Contributor
Author
|
@efriedma-quic @asl Can one of you review this as clang codegen maintainers? (FYI: I know the CI is failing. There are ~200 tests that need updating and I figured I would do the mechanical updates after people have approved to not pollute the diff for review). |
Collaborator
efriedma-quic
left a comment
efriedma-quic
left a comment
There was a problem hiding this comment.
I'll want to skim the "mechanical" update to be sure there aren't any unexpected effects. Otherwise LGTM
Contributor
Author
Just pushed the test update. Some of the OpenMP tests had some unrelated diff for tests that used update_cc_test_checks. I've just left it in for now, but can also pull that diff out into a separate PR/omit it altogether if preferred. |
| // CHECK-NEXT: %[[FPLOAD:.*]] = load ptr, ptr %[[FPGEP]] | ||
| // X64-NEXT: %[[CALL:.*]] = call noundef ptr %[[FPLOAD]](ptr noundef nonnull align 8 dereferenceable(8) %[[LPTR]], i32 noundef 3) | ||
| // X86-NEXT: %[[CALL:.*]] = call x86_thiscallcc noundef ptr %[[FPLOAD]](ptr noundef nonnull align 4 dereferenceable(4) %[[LPTR]], i32 noundef 3) | ||
| // X64-NEXT: %[[CALL:.*]] = call noundef ptr %[[FPLOAD]](ptr noundef nonnull align 8 dead_on_return(8) dereferenceable(8) %[[LPTR]], i32 noundef 3) |
Collaborator
There was a problem hiding this comment.
This is a little scary; the call here is actually deleting an array.
I think it might actually be safe to assume here that the array has at least one element, but I'm not sure that generalizes. I'd be more comfortable if the code explicitly checked getDtorType() and excluded Dtor_VectorDeleting.
CC @Fznamznon
Contributor
Author
There was a problem hiding this comment.
Good point. Omitted Dtor_VectorDeleting. In the general case we can't know the size of the array statically, so I don't think it makes sense to annotate the vector deleting destructor anyways.
And this would definitely be incorrect for zero-length arrays.
boomanaiden154
added a commit
that referenced
this pull request
Feb 6, 2026
)" This reverts commit 278fd05. This was causing some buildbot failures in DSE: 1. https://lab.llvm.org/buildbot/#/builders/25/builds/15425 2. https://lab.llvm.org/buildbot/#/builders/24/builds/17304
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/51/builds/31550 Here is the relevant piece of the build log for the reference |
boomanaiden154
added a commit
that referenced
this pull request
Feb 8, 2026
rishabhmadan19
pushed a commit
to rishabhmadan19/llvm-project
that referenced
this pull request
Feb 9, 2026
This helps to clean up any dead stores that come up at the end of the destructor. The motivating example was a refactoring in libc++'s basic_string implementation in 8dae17b that added a zeroing store into the destructor, causing a large performance regression on an internal workload. We also saw a ~0.2% performance increase on an internal server workload when enabling this. I also tested this against all of the non-flaky tests in our large C++ codebase and found a minimal number of issues that all happened to be in user code.
rishabhmadan19
pushed a commit
to rishabhmadan19/llvm-project
that referenced
this pull request
Feb 9, 2026
…#166276)" This reverts commit 278fd05. This was causing some buildbot failures in DSE: 1. https://lab.llvm.org/buildbot/#/builders/25/builds/15425 2. https://lab.llvm.org/buildbot/#/builders/24/builds/17304
rishabhmadan19
pushed a commit
to rishabhmadan19/llvm-project
that referenced
this pull request
Feb 9, 2026
Xinlong-Chen
pushed a commit
to Xinlong-Chen/llvm-project
that referenced
this pull request
Feb 12, 2026
This helps to clean up any dead stores that come up at the end of the destructor. The motivating example was a refactoring in libc++'s basic_string implementation in 8dae17b that added a zeroing store into the destructor, causing a large performance regression on an internal workload. We also saw a ~0.2% performance increase on an internal server workload when enabling this. I also tested this against all of the non-flaky tests in our large C++ codebase and found a minimal number of issues that all happened to be in user code.
Xinlong-Chen
pushed a commit
to Xinlong-Chen/llvm-project
that referenced
this pull request
Feb 12, 2026
…#166276)" This reverts commit 278fd05. This was causing some buildbot failures in DSE: 1. https://lab.llvm.org/buildbot/#/builders/25/builds/15425 2. https://lab.llvm.org/buildbot/#/builders/24/builds/17304
Xinlong-Chen
pushed a commit
to Xinlong-Chen/llvm-project
that referenced
this pull request
Feb 12, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
11 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This helps to clean up any dead stores that come up at the end of the destructor. The motivating example was a refactoring in libc++'s basic_string implementation in 8dae17b that added a zeroing store into the destructor, causing a large performance regression on an internal workload. We also saw a ~0.2% performance increase on an internal server workload when enabling this.
I also tested this against all of the non-flaky tests in our large C++ codebase and found a minimal number of issues that all happened to be in user code.