[Windows] Put init_seg(compiler/lib) in llvm.global_ctors

[aeubanks]

Summary

Currently we treat initializers with init_seg(compiler/lib) as similar
to any other init_seg, they simply have a global variable in the proper
section (".CRT$XCC" for compiler/".CRT$XCL" for lib) and are added to
llvm.used. However, this doesn't match with how LLVM sees normal (or
init_seg(user)) initializers via llvm.global_ctors. This
causes issues like incorrect init_seg(compiler) vs init_seg(user)
ordering due to GlobalOpt evaluating constructors, and the
ability to remove init_seg(compiler/lib) initializers at all.

Currently we use 'A' for priorities less than 200. Use 200 for init_seg(compiler) (".CRT$XCC") and 400 for init_seg(lib) (".CRT$XCL"), which do not append the priority to the section name. Priorities between 200 and 400 use ".CRT$XCC${Priority}". This allows for some wiggle room for people/future extensions that want to add initializers between compiler and lib.

Fixes #56922

[rnk]

Please update the AttrDocs.td file to document the relationship between init_seg and init_priority on Windows.

Can we increase the threshold for library so that we can insert initializers between compiler and library in the future? Maybe Microsoft will come out with some new semantic level that goes between C and L.

This is my proposed mapping:

• 101-199: A${N}, for very early things (ASan, I think)
• 200: C, compiler
• 201-399: C${N}, between compiler and library
• 400: L, library
• 401-65534: T${N}, after library, before user
• U: user, 65535

[aeubanks]

use 400 for lib, update documentation

[rnk]

lgtm, thanks!

clang/include/clang/Basic/AttrDocs.td

.. code-block:: c++
 
   struct SomeType { SomeType(); };
   __attribute__((init_priority(200))) SomeType Obj1;
   __attribute__((init_priority(101))) SomeType Obj2;
 
 ``Obj2`` will be initialized *before* ``Obj1`` despite the usual order of
 initialization being the opposite.
 
+On Windows, ``init_seg(compiler)`` is represented with a priority of 200 and
+``init_seg(library)`` is represented with a priority of 400. ``init_seg(user)``
+uses the default 65535 priority.
+
 This attribute is only supported for C++ and Objective-C++ and is ignored in
 other language modes. Currently, this attribute is not implemented on z/OS.
   }];
 }
 
 def InitSegDocs : Documentation {
   let Category = DocCatVariable;
   let Content = [{

clang/lib/CodeGen/CGDeclCXX.cpp

CodeGenModule::EmitCXXGlobalVarDeclInitFunc(const VarDecl *D,
 
   if (D->getTLSKind()) {
     // FIXME: Should we support init_priority for thread_local?
     // FIXME: We only need to register one __cxa_thread_atexit function for the
     // entire TU.
     CXXThreadLocalInits.push_back(Fn);
     CXXThreadLocalInitVars.push_back(D);
   } else if (PerformInit && ISA) {
+    // Contract with backend that "init_seg(compiler)" corresponds to priority
+    // 200 and "init_seg(lib)" corresponds to priority 400.
+    int Priority = -1;
+    if (ISA->getSection() == ".CRT$XCC")
+      Priority = 200;
+    else if (ISA->getSection() == ".CRT$XCL")
+      Priority = 400;
+
+    if (Priority != -1)
+      AddGlobalCtor(Fn, Priority, COMDATKey);
+    else
       EmitPointerToInitFunc(D, Addr, Fn, ISA);
   } else if (auto *IPA = D->getAttr<InitPriorityAttr>()) {
     OrderGlobalInitsOrStermFinalizers Key(IPA->getPriority(),
                                           PrioritizedCXXGlobalInits.size());
     PrioritizedCXXGlobalInits.push_back(std::make_pair(Key, Fn));
   } else if (isTemplateInstantiation(D->getTemplateSpecializationKind()) ||
              getContext().GetGVALinkageForVariable(D) == GVA_DiscardableODR ||
              D->hasAttr<SelectAnyAttr>()) {
     // C++ [basic.start.init]p2:

clang/test/CodeGenCXX/pragma-init_seg.cpp

 // RUN: %clang_cc1 %s -triple=i686-pc-win32 -fms-extensions -emit-llvm -o - | FileCheck %s
 
 int f();
 
 // CHECK: $"?x@selectany_init@@3HA" = comdat any
 // CHECK: $"?x@?$A@H@explicit_template_instantiation@@2HB" = comdat any
 // CHECK: $"?x@?$A@H@implicit_template_instantiation@@2HB" = comdat any
 
 namespace simple_init {
 #pragma init_seg(compiler)
 int x = f();
 // CHECK: @"?x@simple_init@@3HA" = dso_local global i32 0, align 4
-// CHECK: @__cxx_init_fn_ptr = private constant ptr @"??__Ex@simple_init@@YAXXZ", section ".CRT$XCC"
+// No function pointer!  This one goes on @llvm.global_ctors.
 
 #pragma init_seg(lib)
 int y = f();
 // CHECK: @"?y@simple_init@@3HA" = dso_local global i32 0, align 4
-// CHECK: @__cxx_init_fn_ptr.1 = private constant ptr @"??__Ey@simple_init@@YAXXZ", section ".CRT$XCL"
+// No function pointer!  This one goes on @llvm.global_ctors.
 
 #pragma init_seg(user)
 int z = f();
 // CHECK: @"?z@simple_init@@3HA" = dso_local global i32 0, align 4
 // No function pointer!  This one goes on @llvm.global_ctors.
 }
 
 #pragma init_seg(".asdf")
 
 namespace internal_init {
 namespace {
 int x = f();
 // CHECK: @"?x@?A0x{{[^@]*}}@internal_init@@3HA" = internal global i32 0, align 4
-// CHECK: @__cxx_init_fn_ptr.2 = private constant ptr @"??__Ex@?A0x{{[^@]*}}@internal_init@@YAXXZ", section ".asdf"
+// CHECK: @__cxx_init_fn_ptr = private constant ptr @"??__Ex@?A0x{{[^@]*}}@internal_init@@YAXXZ", section ".asdf"
 }
 }
 
 namespace selectany_init {
 int __declspec(selectany) x = f();
 // CHECK: @"?x@selectany_init@@3HA" = weak_odr dso_local global i32 0, comdat, align 4
-// CHECK: @__cxx_init_fn_ptr.3 = private constant ptr @"??__Ex@selectany_init@@YAXXZ", section ".asdf", comdat($"?x@selectany_init@@3HA")
+// CHECK: @__cxx_init_fn_ptr.1 = private constant ptr @"??__Ex@selectany_init@@YAXXZ", section ".asdf", comdat($"?x@selectany_init@@3HA")
 }
 
 namespace explicit_template_instantiation {
 template <typename T> struct A { static const int x; };
 template <typename T> const int A<T>::x = f();
 template struct A<int>;
 // CHECK: @"?x@?$A@H@explicit_template_instantiation@@2HB" = weak_odr dso_local global i32 0, comdat, align 4
-// CHECK: @__cxx_init_fn_ptr.4 = private constant ptr @"??__E?x@?$A@H@explicit_template_instantiation@@2HB@@YAXXZ", section ".asdf", comdat($"?x@?$A@H@explicit_template_instantiation@@2HB")
+// CHECK: @__cxx_init_fn_ptr.2 = private constant ptr @"??__E?x@?$A@H@explicit_template_instantiation@@2HB@@YAXXZ", section ".asdf", comdat($"?x@?$A@H@explicit_template_instantiation@@2HB")
 }
 
 namespace implicit_template_instantiation {
 template <typename T> struct A { static const int x; };
 template <typename T> const int A<T>::x = f();
 int g() { return A<int>::x; }
 // CHECK: @"?x@?$A@H@implicit_template_instantiation@@2HB" = linkonce_odr dso_local global i32 0, comdat, align 4
-// CHECK: @__cxx_init_fn_ptr.5 = private constant ptr @"??__E?x@?$A@H@implicit_template_instantiation@@2HB@@YAXXZ", section ".asdf", comdat($"?x@?$A@H@implicit_template_instantiation@@2HB")
+// CHECK: @__cxx_init_fn_ptr.3 = private constant ptr @"??__E?x@?$A@H@implicit_template_instantiation@@2HB@@YAXXZ", section ".asdf", comdat($"?x@?$A@H@implicit_template_instantiation@@2HB")
 }
 
 // ... and here's where we emitted user level ctors.
-// CHECK: @llvm.global_ctors = appending global [1 x { i32, ptr, ptr }]
-// CHECK: [{ i32, ptr, ptr } { i32 65535, ptr @_GLOBAL__sub_I_pragma_init_seg.cpp, ptr null }]
+// CHECK: @llvm.global_ctors = appending global [3 x { i32, ptr, ptr }]
+// CHECK: [{ i32, ptr, ptr } { i32 200, ptr @"??__Ex@simple_init@@YAXXZ", ptr @"?x@simple_init@@3HA" }, { i32, ptr, ptr } { i32 400, ptr @"??__Ey@simple_init@@YAXXZ", ptr @"?y@simple_init@@3HA" }, { i32, ptr, ptr } { i32 65535, ptr @_GLOBAL__sub_I_pragma_init_seg.cpp, ptr null }]
 
 // We have to mark everything used so we can survive globalopt, even through
 // LTO.  There's no way LLVM could really understand if data in the .asdf
 // section is really used or dead.
 //
-// CHECK: @llvm.used = appending global [6 x ptr]
+// CHECK: @llvm.used = appending global [4 x ptr]
 // CHECK: [ptr @__cxx_init_fn_ptr,
 // CHECK: ptr @__cxx_init_fn_ptr.1,
 // CHECK: ptr @__cxx_init_fn_ptr.2,
-// CHECK: ptr @__cxx_init_fn_ptr.3,
-// CHECK: ptr @__cxx_init_fn_ptr.4,
-// CHECK: ptr @__cxx_init_fn_ptr.5], section "llvm.metadata"
+// CHECK: ptr @__cxx_init_fn_ptr.3], section "llvm.metadata"

llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp

if (T.isWindowsMSVCEnvironment() || T.isWindowsItaniumEnvironment()) {
     if (Priority == 65535)
       return Ctx.getAssociativeCOFFSection(Default, KeySym, 0);
 
     // Otherwise, we need to compute a new section name. Low priorities should
     // run earlier. The linker will sort sections ASCII-betically, and we need a
     // string that sorts between .CRT$XCA and .CRT$XCU. In the general case, we
     // make a name like ".CRT$XCT12345", since that runs before .CRT$XCU. Really
     // low priorities need to sort before 'L', since the CRT uses that
-    // internally, so we use ".CRT$XCA00001" for them.
+    // internally, so we use ".CRT$XCA00001" for them. We have a contract with
+    // the frontend that "init_seg(compiler)" corresponds to priority 200 and
+    // "init_seg(lib)" corresponds to priority 400, and those respectively use
+    // 'C' and 'L' without the priority suffix. Priorities between 200 and 400
+    // use 'C' with the priority as a suffix.
     SmallString<24> Name;
+    char LastLetter = 'T';
+    bool AddPrioritySuffix = Priority != 200 && Priority != 400;
+    if (Priority < 200)
+      LastLetter = 'A';
+    else if (Priority < 400)
+      LastLetter = 'C';
+    else if (Priority == 400)
+      LastLetter = 'L';
     raw_svector_ostream OS(Name);
-    OS << ".CRT$X" << (IsCtor ? "C" : "T") <<
-        (Priority < 200 ? 'A' : 'T') << format("%05u", Priority);
+    OS << ".CRT$X" << (IsCtor ? "C" : "T") << LastLetter;
+    if (AddPrioritySuffix)
+      OS << format("%05u", Priority);
     MCSectionCOFF *Sec = Ctx.getCOFFSection(
         Name, COFF::IMAGE_SCN_CNT_INITIALIZED_DATA | COFF::IMAGE_SCN_MEM_READ,
         SectionKind::getReadOnly());
     return Ctx.getAssociativeCOFFSection(Sec, KeySym, 0);
   }
 
   std::string Name = IsCtor ? ".ctors" : ".dtors";
   if (Priority != 65535)

llvm/test/CodeGen/X86/ctor-priority-coff.ll

 ; RUN: llc < %s | FileCheck %s
 
 ; Check that we come up with appropriate section names that link.exe sorts
 ; well.
 
 ; CHECK: .section        .CRT$XCA00042,"dr"
 ; CHECK: .p2align        3
 ; CHECK: .quad   f
+; CHECK: .section        .CRT$XCC,"dr"
+; CHECK: .p2align        3
+; CHECK: .quad   i
+; CHECK: .section        .CRT$XCC00250,"dr"
+; CHECK: .p2align        3
+; CHECK: .quad   k
+; CHECK: .section        .CRT$XCL,"dr"
+; CHECK: .p2align        3
+; CHECK: .quad   j
 ; CHECK: .section        .CRT$XCT12345,"dr"
 ; CHECK: .p2align        3
 ; CHECK: .quad   g
 ; CHECK: .section        .CRT$XCT23456,"dr",associative,h
 ; CHECK: .p2align        3
 ; CHECK: .quad   init_h
 
 target datalayout = "e-m:w-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-pc-windows-msvc19.14.26433"
 
 $h = comdat any
 
 @h = linkonce_odr global i8 55, comdat, align 1
 
 @str0 = private dso_local unnamed_addr constant [6 x i8] c"later\00", align 1
 @str1 = private dso_local unnamed_addr constant [6 x i8] c"first\00", align 1
 @str2 = private dso_local unnamed_addr constant [5 x i8] c"main\00", align 1
 
[e-mail]_ctors = appending global [3 x { i32, ptr, ptr }] [
[e-mail]_ctors = appending global [6 x { i32, ptr, ptr }] [
   { i32, ptr, ptr } { i32 12345, ptr @g, ptr null },
   { i32, ptr, ptr } { i32 42, ptr @f, ptr null },
-  { i32, ptr, ptr } { i32 23456, ptr @init_h, ptr @h }
+  { i32, ptr, ptr } { i32 23456, ptr @init_h, ptr @h },
+  { i32, ptr, ptr } { i32 200, ptr @i, ptr null },
+  { i32, ptr, ptr } { i32 400, ptr @j, ptr null },
+  { i32, ptr, ptr } { i32 250, ptr @k, ptr null }
 ]
 
 declare dso_local i32 @puts(ptr nocapture readonly) local_unnamed_addr
 
 define dso_local void @g() {
 entry:
   %call = tail call i32 @puts(ptr @str0)
   ret void
 }
 
 define dso_local void @f() {
 entry:
   %call = tail call i32 @puts(ptr @str1)
   ret void
 }
 
 define dso_local void @init_h() {
 entry:
   store i8 42, ptr @h
   ret void
 }
 
+define dso_local void @i() {
+entry:
+  store i8 43, ptr @h
+  ret void
+}
+
+define dso_local void @j() {
+entry:
+  store i8 44, ptr @h
+  ret void
+}
+
+define dso_local void @k() {
+entry:
+  store i8 45, ptr @h
+  ret void
+}
 
 ; Function Attrs: nounwind uwtable
 define dso_local i32 @main() local_unnamed_addr {
 entry:
   %call = tail call i32 @puts(ptr @str2)
   ret i32 0
 }