[MC] Set sh_link to 0 if the associated symbol is undefined

Part of https://bugs.llvm.org/show_bug.cgi?id=41734

LTO can drop externally available definitions. Such AssociatedSymbol is
not associated with a symbol. ELFWriter::writeSection() will assert.

Allow a SHF_LINK_ORDER section to have sh_link=0.

We need to give sh_link a syntax, a literal zero in the linked-to symbol
position, e.g. `.section name,"ao",@progbits,0`

Reviewed By: pcc

Differential Revision: https://reviews.llvm.org/D72899

diff --git a/llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp b/llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp
index 4859919..8ef9125 100644 (file)
--- a/llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp
+++ b/llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp
@@ -680,7 +680,7 @@ MCSection *TargetLoweringObjectFileELF::getExplicitSectionGlobal(
   // MD_associated in a unique section.
   unsigned UniqueID = MCContext::GenericSectionID;
   const MCSymbolELF *LinkedToSym = getLinkedToSymbol(GO, TM);
-  if (LinkedToSym) {
+  if (GO->getMetadata(LLVMContext::MD_associated)) {
     UniqueID = NextUniqueID++;
     Flags |= ELF::SHF_LINK_ORDER;
   } else {

diff --git a/llvm/lib/MC/ELFObjectWriter.cpp b/llvm/lib/MC/ELFObjectWriter.cpp
index 5a5692c..b44a36b 100644 (file)
--- a/llvm/lib/MC/ELFObjectWriter.cpp
+++ b/llvm/lib/MC/ELFObjectWriter.cpp
@@ -1024,9 +1024,13 @@ void ELFWriter::writeSection(const SectionIndexMapTy &SectionIndexMap,
   }
 
   if (Section.getFlags() & ELF::SHF_LINK_ORDER) {
+    // If the value in the associated metadata is not a definition, Sym will be
+    // undefined. Represent this with sh_link=0.
     const MCSymbol *Sym = Section.getLinkedToSymbol();
-    const MCSectionELF *Sec = cast<MCSectionELF>(&Sym->getSection());
-    sh_link = SectionIndexMap.lookup(Sec);
+    if (Sym && Sym->isInSection()) {
+      const MCSectionELF *Sec = cast<MCSectionELF>(&Sym->getSection());
+      sh_link = SectionIndexMap.lookup(Sec);
+    }
   }
 
   WriteSecHdrEntry(StrTabBuilder.getOffset(Section.getName()),

diff --git a/llvm/lib/MC/MCParser/ELFAsmParser.cpp b/llvm/lib/MC/MCParser/ELFAsmParser.cpp
index e5ab13b..41779d0 100644 (file)
--- a/llvm/lib/MC/MCParser/ELFAsmParser.cpp
+++ b/llvm/lib/MC/MCParser/ELFAsmParser.cpp
@@ -450,8 +450,14 @@ bool ELFAsmParser::parseLinkedToSym(MCSymbolELF *&LinkedToSym) {
   Lex();
   StringRef Name;
   SMLoc StartLoc = L.getLoc();
-  if (getParser().parseIdentifier(Name))
+  if (getParser().parseIdentifier(Name)) {
+    if (getParser().getTok().getString() == "0") {
+      getParser().Lex();
+      LinkedToSym = nullptr;
+      return false;
+    }
     return TokError("invalid linked-to symbol");
+  }
   LinkedToSym = dyn_cast_or_null<MCSymbolELF>(getContext().lookupSymbol(Name));
   if (!LinkedToSym || !LinkedToSym->isInSection())
     return Error(StartLoc, "linked-to symbol is not in a section: " + Name);

diff --git a/llvm/lib/MC/MCSectionELF.cpp b/llvm/lib/MC/MCSectionELF.cpp
index 77c259c..7a15556 100644 (file)
--- a/llvm/lib/MC/MCSectionELF.cpp
+++ b/llvm/lib/MC/MCSectionELF.cpp
@@ -172,9 +172,11 @@ void MCSectionELF::PrintSwitchToSection(const MCAsmInfo &MAI, const Triple &T,
   }
 
   if (Flags & ELF::SHF_LINK_ORDER) {
-    assert(LinkedToSym);
     OS << ",";
-    printName(OS, LinkedToSym->getName());
+    if (LinkedToSym)
+      printName(OS, LinkedToSym->getName());
+    else
+      OS << '0';
   }
 
   if (isUnique())

diff --git a/llvm/test/CodeGen/X86/elf-associated-discarded.ll b/llvm/test/CodeGen/X86/elf-associated-discarded.ll
new file mode 100644 (file)
index 0000000..5a4fad4
--- /dev/null
+++ b/llvm/test/CodeGen/X86/elf-associated-discarded.ll
@@ -0,0 +1,23 @@
+;; Test that we keep SHF_LINK_ORDER but reset sh_link to 0 if the associated
+;; symbol is not defined.
+; RUN: llc -mtriple=x86_64 -data-sections=1 < %s | FileCheck %s
+; RUN: llc -filetype=obj -mtriple=x86_64 -data-sections=1 < %s | llvm-readelf -S - | FileCheck --check-prefix=SEC %s
+
+;; FIXME The assembly output cannot be assembled because foo is not defined.
+;; This is difficult to fix because we allow loops (see elf-associated.ll
+;; .data.c and .data.d).
+; CHECK: .section .data.a,"awo",@progbits,foo
+; CHECK: .section .data.b,"awo",@progbits,foo
+
+;; No 'L' (SHF_LINK_ORDER). sh_link=0.
+; SEC; Name    {{.*}} Flg Lk Inf
+; SEC: .data.a {{.*}} WAL  0   0
+; SEC: .data.b {{.*}} WAL  0   0
+
+;; The definition may be discarded by LTO.
+declare void @foo()
+
+@a = global i32 1, !associated !0
+@b = global i32 1, !associated !0
+
+!0 = !{void ()* @foo}

diff --git a/llvm/test/CodeGen/X86/elf-associated.ll b/llvm/test/CodeGen/X86/elf-associated.ll
index e0e9e00..14a4b5b 100644 (file)
--- a/llvm/test/CodeGen/X86/elf-associated.ll
+++ b/llvm/test/CodeGen/X86/elf-associated.ll
@@ -36,15 +36,15 @@
 ; Non-GlobalValue metadata.
 @l = global i32 1, section "ccc", !associated !5
 !5 = !{i32* null}
-; CHECK-DAG: .section  ccc,"aw",@progbits
+; CHECK-DAG: .section  ccc,"awo",@progbits,0,unique,3
 
 ; Null metadata.
 @m = global i32 1, section "ddd", !associated !6
 !6 = distinct !{null}
-; CHECK-DAG: .section  ddd,"aw",@progbits
+; CHECK-DAG: .section  ddd,"awo",@progbits,0,unique,4
 
 ; Aliases are OK.
 @n = alias i32, i32* inttoptr (i64 add (i64 ptrtoint (i32* @a to i64), i64 1297036692682702848) to i32*)
 @o = global i32 1, section "eee", !associated !7
 !7 = !{i32* @n}
-; CHECK-DAG: .section  eee,"awo",@progbits,n,unique,3
+; CHECK-DAG: .section  eee,"awo",@progbits,n,unique,5

diff --git a/llvm/test/MC/ELF/section-linkorder.s b/llvm/test/MC/ELF/section-linkorder.s
new file mode 100644 (file)
index 0000000..a0f6357
--- /dev/null
+++ b/llvm/test/MC/ELF/section-linkorder.s
@@ -0,0 +1,8 @@
+# RUN: llvm-mc -triple x86_64 %s | FileCheck %s --check-prefix=ASM
+# RUN: llvm-mc -filetype=obj -triple=x86_64 %s -o %t
+# RUN: llvm-readelf -S %t | FileCheck %s
+
+# ASM:   .section .linkorder,"ao",@progbits,0
+# CHECK: Name       Type     {{.*}} Flg Lk
+# CHECK: .linkorder PROGBITS {{.*}}  AL  0
+.section .linkorder,"ao",@progbits,0