[MemProf] Recognize hot/cold operator new as replaceable allocations

[tejohnson]

Summary

Follow up to LLVM-side change to allow conversion to hot/cold hinted
operator new, not yet standard but supported by open source tcmalloc.
The LLVM change in a35206d78280e0ebcf48cd21bc5fff5c3b9c73fa added the
necessary support to recognize these as library functions, and we
subsequently found that a clang-side change is needed to give them
builtin/nobuiltin attributes as appropriate so they have consistent
removeability.

Based on discussion with Richard Smith, converted the parameter type to
a reserved identifier (39f7b48671dae5fbe3afc49f33f50c2b6340bb89) and
added support in this patch to recognize it in clang.

[snehasish]

lgtm

clang/test/CodeGenCXX/new_hot_cold.cpp

16

Can we skip the typedef and just say enum class __hot_cold_t : unsigned char;?

Also it might be useful to link to the tcmalloc documentation for hot_cold_t to document the expected underlying type?
https://github.com/google/tcmalloc/blob/220043886d4e2efff7a5702d5172cb8065253664/tcmalloc/malloc_extension.h#L53

[tejohnson]

clang/test/CodeGenCXX/new_hot_cold.cpp

16

Can we skip the typedef and just say enum class __hot_cold_t : unsigned char;?

I did that specifically to test the way it is declared in tcmalloc. In particular, this tests the TypedefType while loop at lines 3301-2 in Decl.cpp which was required to "see through" that and handle that case properly. I can add a comment here to the test. Will also add the link to tcmalloc here and in Decl.cpp

[tejohnson]

Address comment

clang/lib/AST/Decl.cpp

bool FunctionDecl::isReplaceableGlobalAllocationFunction(
   if (isa<CXXRecordDecl>(getDeclContext()))
     return false;
 
   // This can only fail for an invalid 'operator new' declaration.
   if (!getDeclContext()->getRedeclContext()->isTranslationUnit())
     return false;
 
   const auto *FPT = getType()->castAs<FunctionProtoType>();
-  if (FPT->getNumParams() == 0 || FPT->getNumParams() > 3 || FPT->isVariadic())
+  if (FPT->getNumParams() == 0 || FPT->getNumParams() > 4 || FPT->isVariadic())
     return false;
 
   // If this is a single-parameter function, it must be a replaceable global
   // allocation or deallocation function.
   if (FPT->getNumParams() == 1)
     return true;
 
   unsigned Params = 1;
bool FunctionDecl::isReplaceableGlobalAllocationFunction(
   // In C++17, the next parameter can be a 'std::align_val_t' for aligned
   // new/delete.
   if (Ctx.getLangOpts().AlignedAllocation && !Ty.isNull() && Ty->isAlignValT()) {
     Consume();
     if (AlignmentParam)
       *AlignmentParam = Params;
   }
 
-  // Finally, if this is not a sized delete, the final parameter can
-  // be a 'const std::nothrow_t&'.
+  // If this is not a sized delete, the next parameter can be a
+  // 'const std::nothrow_t&'.
   if (!IsSizedDelete && !Ty.isNull() && Ty->isReferenceType()) {
     Ty = Ty->getPointeeType();
     if (Ty.getCVRQualifiers() != Qualifiers::Const)
       return false;
     if (Ty->isNothrowT()) {
       if (IsNothrow)
         *IsNothrow = true;
       Consume();
     }
   }
 
+  // Finally, recognize the not yet standard versions of new that take a
+  // hot/cold allocation hint (__hot_cold_t). These are currently supported by
+  // tcmalloc.
+  if (!IsSizedDelete && !Ty.isNull() && Ty->isEnumeralType()) {
+    QualType T = Ty;
+    while (const auto *TD = T->getAs<TypedefType>())
+      T = TD->getDecl()->getUnderlyingType();
+    IdentifierInfo *II = T->getAs<EnumType>()->getDecl()->getIdentifier();
+    if (II && II->isStr("__hot_cold_t"))
+      Consume();
+  }
+
   return Params == FPT->getNumParams();
 }
 
 bool FunctionDecl::isInlineBuiltinDeclaration() const {
   if (!getBuiltinID())
     return false;
 
   const FunctionDecl *Definition;

clang/test/CodeGenCXX/new_hot_cold.cpp

+// Test to check that the appropriate attributes are added to the __hot_cold_t
+// versions of operator new.
+
+// RUN: %clang_cc1 -std=c++14 -triple x86_64-unknown-unknown %s -faligned-allocation -emit-llvm -o - | FileCheck %s
+
+typedef __typeof__(sizeof(0)) size_t;
+
+namespace std {
+struct nothrow_t {};
+enum class align_val_t : size_t;
+} // namespace std
+std::nothrow_t nothrow;
+
+typedef unsigned char uint8_t;
+
+enum class __hot_cold_t : uint8_t;
+namespace malloc_namespace {
+using hot_cold_t = __hot_cold_t;
+}
+
+void *operator new(size_t size,
+                   malloc_namespace::hot_cold_t hot_cold) noexcept(false);
+void *operator new[](size_t size,
+                     malloc_namespace::hot_cold_t hot_cold) noexcept(false);
+void *operator new(size_t size, const std::nothrow_t &,
+                   malloc_namespace::hot_cold_t hot_cold) noexcept;
+void *operator new[](size_t size, const std::nothrow_t &,
+                     malloc_namespace::hot_cold_t hot_cold) noexcept;
+void *operator new(size_t size, std::align_val_t alignment,
+                   malloc_namespace::hot_cold_t hot_cold) noexcept(false);
+void *operator new[](size_t size, std::align_val_t alignment,
+                     malloc_namespace::hot_cold_t hot_cold) noexcept(false);
+void *operator new(size_t size, std::align_val_t alignment,
+                   const std::nothrow_t &,
+                   malloc_namespace::hot_cold_t hot_cold) noexcept;
+void *operator new[](size_t size, std::align_val_t alignment,
+                     const std::nothrow_t &,
+                     malloc_namespace::hot_cold_t hot_cold) noexcept;
+
+// All explicit operator new calls should not get any builtin attribute, whereas
+// all implicit new expressions should get builtin attributes. All of the
+// declarations should get nobuiltin attributes.
+
+void hot_cold_new() {
+  // CHECK: call noalias noundef nonnull ptr @_Znwm12__hot_cold_t(i64 noundef 1, i8 noundef zeroext 0) [[ATTR_NO_BUILTIN_CALL:#[^ ]*]]
+  operator new(1, (__hot_cold_t)0);
+  // CHECK: call noalias noundef nonnull ptr @_Znwm12__hot_cold_t(i64 noundef 4, i8 noundef zeroext 0) [[ATTR_BUILTIN_CALL:#[^ ]*]]
+  new ((__hot_cold_t)0) int;
+}
+
+// CHECK: declare noundef nonnull ptr @_Znwm12__hot_cold_t(i64 noundef, i8 noundef zeroext) [[ATTR_NOBUILTIN:#[^ ]*]]
+
+void hot_cold_new_array() {
+  // CHECK: call noalias noundef nonnull ptr @_Znam12__hot_cold_t(i64 noundef 1, i8 noundef zeroext 0) [[ATTR_NO_BUILTIN_CALL:#[^ ]*]]
+  operator new[](1, (__hot_cold_t)0);
+  // CHECK: call noalias noundef nonnull ptr @_Znam12__hot_cold_t(i64 noundef 4, i8 noundef zeroext 0) [[ATTR_BUILTIN_CALL:#[^ ]*]]
+  new ((__hot_cold_t)0) int[1];
+}
+
+// CHECK: declare noundef nonnull ptr @_Znam12__hot_cold_t(i64 noundef, i8 noundef zeroext) [[ATTR_NOBUILTIN:#[^ ]*]]
+
+void hot_cold_new_nothrow() {
+  // CHECK: call noalias noundef ptr @_ZnwmRKSt9nothrow_t12__hot_cold_t(i64 noundef 1, {{.*}} @nothrow, i8 noundef zeroext 0) [[ATTR_NO_BUILTIN_NOTHROW_CALL:#[^ ]*]]
+  operator new(1, nothrow, (__hot_cold_t)0);
+  // CHECK: call noalias noundef ptr @_ZnwmRKSt9nothrow_t12__hot_cold_t(i64 noundef 4, {{.*}} @nothrow, i8 noundef zeroext 0) [[ATTR_BUILTIN_NOTHROW_CALL:#[^ ]*]]
+  new (nothrow, (__hot_cold_t)0) int;
+}
+
+// CHECK: declare noundef ptr @_ZnwmRKSt9nothrow_t12__hot_cold_t(i64 noundef, ptr noundef nonnull align 1 dereferenceable(1), i8 noundef zeroext) [[ATTR_NOBUILTIN_NOTHROW:#[^ ]*]]
+
+void hot_cold_new_nothrow_array() {
+  // CHECK: call noalias noundef ptr @_ZnamRKSt9nothrow_t12__hot_cold_t(i64 noundef 1, {{.*}} @nothrow, i8 noundef zeroext 0) [[ATTR_NO_BUILTIN_NOTHROW_CALL:#[^ ]*]]
+  operator new[](1, nothrow, (__hot_cold_t)0);
+  // CHECK: call noalias noundef ptr @_ZnamRKSt9nothrow_t12__hot_cold_t(i64 noundef 4, {{.*}} @nothrow, i8 noundef zeroext 0) [[ATTR_BUILTIN_NOTHROW_CALL:#[^ ]*]]
+  new (nothrow, (__hot_cold_t)0) int[1];
+}
+
+// CHECK: declare noundef ptr @_ZnamRKSt9nothrow_t12__hot_cold_t(i64 noundef, ptr noundef nonnull align 1 dereferenceable(1), i8 noundef zeroext) [[ATTR_NOBUILTIN_NOTHROW:#[^ ]*]]
+
+class alignas(32) alignedstruct {
+  int x;
+};
+
+void hot_cold_new_align() {
+  // CHECK: call noalias noundef nonnull align 32 ptr @_ZnwmSt11align_val_t12__hot_cold_t(i64 noundef 1, i64 noundef 32, i8 noundef zeroext 0) [[ATTR_NO_BUILTIN_CALL:#[^ ]*]]
+  operator new(1, (std::align_val_t)32, (__hot_cold_t)0);
+  // CHECK: call noalias noundef nonnull align 32 ptr @_ZnwmSt11align_val_t12__hot_cold_t(i64 noundef 32, i64 noundef 32, i8 noundef zeroext 0) [[ATTR_BUILTIN_CALL:#[^ ]*]]
+  new ((__hot_cold_t)0) alignedstruct;
+}
+
+// CHECK: declare noundef nonnull ptr @_ZnwmSt11align_val_t12__hot_cold_t(i64 noundef, i64 noundef, i8 noundef zeroext) [[ATTR_NOBUILTIN:#[^ ]*]]
+
+void hot_cold_new_align_array() {
+  // CHECK: call noalias noundef nonnull align 32 ptr @_ZnamSt11align_val_t12__hot_cold_t(i64 noundef 1, i64 noundef 32, i8 noundef zeroext 0) [[ATTR_NO_BUILTIN_CALL:#[^ ]*]]
+  operator new[](1, (std::align_val_t)32, (__hot_cold_t)0);
+  // CHECK: call noalias noundef nonnull align 32 ptr @_ZnamSt11align_val_t12__hot_cold_t(i64 noundef 32, i64 noundef 32, i8 noundef zeroext 0) [[ATTR_BUILTIN_CALL:#[^ ]*]]
+  new ((__hot_cold_t)0) alignedstruct[1];
+}
+
+// CHECK: declare noundef nonnull ptr @_ZnamSt11align_val_t12__hot_cold_t(i64 noundef, i64 noundef, i8 noundef zeroext) [[ATTR_NOBUILTIN:#[^ ]*]]
+
+void hot_cold_new_align_nothrow() {
+  // CHECK: call noalias noundef align 32 ptr @_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t(i64 noundef 1, i64 noundef 32, {{.*}} @nothrow, i8 noundef zeroext 0) [[ATTR_NO_BUILTIN_NOTHROW_CALL:#[^ ]*]]
+  operator new(1, (std::align_val_t)32, nothrow, (__hot_cold_t)0);
+  // CHECK: call noalias noundef align 32 ptr @_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t(i64 noundef 32, i64 noundef 32, {{.*}} @nothrow, i8 noundef zeroext 0) [[ATTR_BUILTIN_NOTHROW_CALL:#[^ ]*]]
+  new (nothrow, (__hot_cold_t)0) alignedstruct;
+}
+
+// CHECK: declare noundef ptr @_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t(i64 noundef, i64 noundef, ptr noundef nonnull align 1 dereferenceable(1), i8 noundef zeroext) [[ATTR_NOBUILTIN_NOTHROW:#[^ ]*]]
+
+void hot_cold_new_align_nothrow_array() {
+  // CHECK: call noalias noundef align 32 ptr @_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t(i64 noundef 1, i64 noundef 32, {{.*}} @nothrow, i8 noundef zeroext 0) [[ATTR_NO_BUILTIN_NOTHROW_CALL:#[^ ]*]]
+  operator new[](1, (std::align_val_t)32, nothrow, (__hot_cold_t)0);
+  // CHECK: call noalias noundef align 32 ptr @_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t(i64 noundef 32, i64 noundef 32, {{.*}} @nothrow, i8 noundef zeroext 0) [[ATTR_BUILTIN_NOTHROW_CALL:#[^ ]*]]
+  new (nothrow, (__hot_cold_t)0) alignedstruct[1];
+}
+
+// CHECK: declare noundef ptr @_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t(i64 noundef, i64 noundef, ptr noundef nonnull align 1 dereferenceable(1), i8 noundef zeroext) [[ATTR_NOBUILTIN_NOTHROW:#[^ ]*]]
+
+// CHECK-DAG: attributes [[ATTR_NOBUILTIN]] = { nobuiltin allocsize(0) {{.*}} }
+// CHECK-DAG: attributes [[ATTR_NOBUILTIN_NOTHROW]] = { nobuiltin nounwind allocsize(0) {{.*}} }
+// CHECK-DAG: attributes [[ATTR_NO_BUILTIN_CALL]] = { allocsize(0) }
+// CHECK-DAG: attributes [[ATTR_BUILTIN_CALL]] = { builtin allocsize(0) }
+// CHECK-DAG: attributes [[ATTR_NO_BUILTIN_NOTHROW_CALL]] = { nounwind allocsize(0) }
+// CHECK-DAG: attributes [[ATTR_BUILTIN_NOTHROW_CALL]] = { builtin nounwind allocsize(0) }