[ThinLTO] Don't create a comdat group for a dropped def with initializer

[tejohnson]

Summary

Non-prevailing weak/linkonce odr symbols will be dropped to
available_externally when possible. If they had an initializer in the
global_ctors list, a comdat group was being created. This code
already had logic to skip available_externally defs, but now the
EliminateAvailableExternally pass will drop these symbols to
declarations earlier. Change the check to skip all declarations for
linker (which includes available_externally along with declarations).

[tejohnson]

Add llvm-lto and llvm-lto2 based tests

[mehdi_amini]

lib/CodeGen/AsmPrinter/AsmPrinter.cpp

1653

Do we need a ThinLTO importing to expose this case?
Can't you have a "simple" llc test that test the CodeGen for this?

[tejohnson]

Replace ThinLTO-specific tests with a simple llc based test

[tejohnson]

lib/CodeGen/AsmPrinter/AsmPrinter.cpp

1653

Good idea, done.

[mehdi_amini]

LGTM.

lib/CodeGen/AsmPrinter/AsmPrinter.cpp

void AsmPrinter::EmitXXStructorList(const DataLayout &DL, const Constant *List,
   unsigned Align = Log2_32(DL.getPointerPrefAlignment());
   std::stable_sort(Structors.begin(), Structors.end(),
                    [](const Structor &L,
                       const Structor &R) { return L.Priority < R.Priority; });
   for (Structor &S : Structors) {
     const TargetLoweringObjectFile &Obj = getObjFileLowering();
     const MCSymbol *KeySym = nullptr;
     if (GlobalValue *GV = S.ComdatKey) {
-      if (GV->hasAvailableExternallyLinkage())
-        // If the associated variable is available_externally, some other TU
+      if (GV->isDeclarationForLinker())
+        // If the associated variable is not defined in this module
+        // (it might be available_externally, or have been an
+        // available_externally definition that was dropped by the
+        // EliminateAvailableExternally pass), some other TU
         // will provide its dynamic initializer.
         continue;
 
       KeySym = getSymbol(GV);
     }
     MCSection *OutputSection =
         (isCtor ? Obj.getStaticCtorSection(S.Priority, KeySym)
                 : Obj.getStaticDtorSection(S.Priority, KeySym));
     OutStreamer->SwitchSection(OutputSection);
     if (OutStreamer->getCurrentSection() != OutStreamer->getPreviousSection())

test/ThinLTO/X86/Inputs/linkonce_comdat.ll

+; ModuleID = 'thinlto_linkonce_comdat1.o'
+source_filename = "thinlto_linkonce_comdat1.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+@fv = linkonce_odr global i8 0, align 8
+
[e-mail]_ctors = appending global [1 x { i32, void ()*, i8* }] [{ i32, void ()*, i8* } { i32 65535, void ()* @__cxx_global_var_init.33, i8* @fv }]
+
+; Function Attrs: nounwind uwtable
+define internal void @__cxx_global_var_init.33() section ".text.startup" {
+  store i8 1, i8* @fv, align 8
+  ret void
+}
+
+; Function Attrs: norecurse nounwind readnone uwtable
+define void @test1() local_unnamed_addr {
+entry:
+  ret void
+}

test/ThinLTO/X86/linkonce_comdat.ll

+; Test to ensure that a global constructor and associated comdat isn't
+; created when ThinLTO has dropped the non-prevailing copy of a
+; linkonce_odr (otherwise we will end up with an incomplete comdat
+; group).
+
+; Do setup work for all below tests: generate bitcode and combined index
+; RUN: opt -module-summary %s -o %t1.bc
+; RUN: opt -module-summary %p/Inputs/linkonce_comdat.ll -o %t2.bc
+; RUN: llvm-lto -thinlto-action=thinlink -o %t3.bc %t1.bc %t2.bc
+
+; RUN: llvm-lto -thinlto-action=run -exported-symbol=main -exported-symbol=fv %t1.bc %t2.bc
+; RUN: llvm-readobj -elf-section-groups %t1.bc.thinlto.o | FileCheck  %s --check-prefix=OBJ1
+; RUN: llvm-readobj -elf-section-groups %t2.bc.thinlto.o | FileCheck  %s --check-prefix=OBJ2
+
+; RUN: llvm-lto2 %t1.bc %t2.bc -o %t.o -save-temps \
+; RUN:     -r=%t1.bc,fv,plx \
+; RUN:     -r=%t1.bc,test1,l \
+; RUN:     -r=%t1.bc,main,plx \
+; RUN:     -r=%t2.bc,fv,lx \
+; RUN:     -r=%t2.bc,test1,pl
+; RUN: llvm-readobj -elf-section-groups %t.o.0 | FileCheck  %s --check-prefix=OBJ1
+; RUN: llvm-readobj -elf-section-groups %t.o.1 | FileCheck  %s --check-prefix=OBJ2
+
+; The prevailing copy in %t1 should result in a complete comdat group.
+; OBJ1:   Group {
+; OBJ1:     Signature: fv
+; OBJ1:     Section(s) in group [
+; OBJ1:       .init_array
+; OBJ1:       .rela.init_array
+
+; We should not generate a comdat group in the object for %t2, even
+; though it originally had an entry in the global_ctors list that
+; would result in creation of a comdat group (unless those that become
+; declarations are not properly ignored).
+; OBJ2: There are no group sections in the file.
+
+; ModuleID = 'thinlto_linkonce_comdat.o'
+source_filename = "thinlto_linkonce_comdat.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+@fv = linkonce_odr global i8 0, align 8
+
[e-mail]_ctors = appending global [1 x { i32, void ()*, i8* }] [{ i32, void ()*, i8* } { i32 65535, void ()* @__cxx_global_var_init.34, i8* @fv }]
+
+define internal void @__cxx_global_var_init.34() section ".text.startup" {
+  store i8 1, i8* @fv, align 8
+  ret void
+}
+
+declare void @test1(...) local_unnamed_addr
+
+; Function Attrs: nounwind uwtable
+define i32 @main() local_unnamed_addr {
+entry:
+  tail call void (...) @test1()
+  ret i32 0
+}

test/tools/gold/X86/Inputs/thinlto_linkonce_comdat1.ll

+; ModuleID = 'thinlto_linkonce_comdat1.c'
+source_filename = "thinlto_linkonce_comdat1.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+@fv = linkonce_odr global i8 0, align 8
+
[e-mail]_ctors = appending global [1 x { i32, void ()*, i8* }] [{ i32, void ()*, i8* } { i32 65535, void ()* @__cxx_global_var_init.33, i8* @fv }]
+
+; Function Attrs: nounwind uwtable
+define internal void @__cxx_global_var_init.33() section ".text.startup" {
+  store i8 1, i8* @fv, align 8
+  ret void
+}
+
+; Function Attrs: norecurse nounwind readnone uwtable
+define void @test1() local_unnamed_addr {
+entry:
+  ret void
+}

test/tools/gold/X86/Inputs/thinlto_linkonce_comdat2.ll

+; ModuleID = 'thinlto_linkonce_comdat2.c'
+source_filename = "thinlto_linkonce_comdat2.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+@fv = linkonce_odr global i8 0, align 8
+
[e-mail]_ctors = appending global [1 x { i32, void ()*, i8* }] [{ i32, void ()*, i8* } { i32 65535, void ()* @__cxx_global_var_init.34, i8* @fv }]
+
+define internal void @__cxx_global_var_init.34() section ".text.startup" {
+  store i8 1, i8* @fv, align 8
+  ret void
+}
+
+; Function Attrs: nounwind uwtable
+define void @test2() local_unnamed_addr {
+entry:
+  tail call void (...) @test1()
+  ret void
+}
+
+declare void @test1(...) local_unnamed_addr

test/tools/gold/X86/thinlto_linkonce_comdat.ll

+; Test to ensure that a global constructor and associated comdat isn't
+; created when ThinLTO has dropped the non-prevailing copy of a
+; linkonce_odr (otherwise we will end up with an incomplete comdat
+; group, which can override the intended complete comdat in the prevailing
+; module).
+
+; First generate bitcode with a module summary index for each file
+; RUN: opt -module-summary %s -o %t.o
+; RUN: opt -module-summary %p/Inputs/thinlto_linkonce_comdat1.ll -o %t2.o
+; RUN: opt -module-summary %p/Inputs/thinlto_linkonce_comdat2.ll -o %t3.o
+
+; Although the objects are ordered "%t2.o %t3.o" in the library, the
+; linker selects %t3.o first since it satisfies a strong reference from
+; %t.o. It later selects %t2.o based on the strong ref from %t3.o.
+; Therefore, %t3.o's copy of @f is prevailing.
+; RUN: %gold -plugin %llvmshlibdir/LLVMgold.so \
+; RUN:    --plugin-opt=thinlto \
+; RUN:    --plugin-opt=save-temps \
+; RUN:    -m elf_x86_64 \
+; RUN:    -o %t4 \
+; RUN:    %t.o \
+; RUN:    --start-lib %t2.o %t3.o --end-lib
+
+; We should not generate a comdat group in the object for %t2, even
+; though it originally had an entry in the global_ctors list that
+; would result in creation of a comdat group (unless those that become
+; declarations are not properly ignored).
+; RUN: llvm-readobj -elf-section-groups %t4.o1 | FileCheck %s --check-prefix=OBJ1
+; OBJ1: There are no group sections in the file.
+
+; The prevailing copy in %t3 should result in a complete comdat group.
+; RUN: llvm-readobj -elf-section-groups %t4.o2 | FileCheck %s --check-prefix=OBJ2
+; OBJ2:   Group {
+; OBJ2:     Signature: fv
+; OBJ2:     Section(s) in group [
+; OBJ2:       .init_array
+; OBJ2:       .rela.init_array
+
+; ModuleID = 'thinlto_linkonce_comdat.c'
+source_filename = "thinlto_linkonce_comdat.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; Function Attrs: nounwind uwtable
+define i32 @main() local_unnamed_addr {
+entry:
+  tail call void (...) @test2()
+  ret i32 0
+}
+
+declare void @test2(...) local_unnamed_addr