[PowerPC] Add .ref in backend for AIX XCOFF to support `-bcdtors:csect` linker option

[tingwang]

Summary

On AIX -bcdtors:csect linker option requires ".ref" pseudo-op to indicate the connection among static variables (static global variable, static class member etc.) and static init/term functions, so that linker can generate the correct output.

Here in IR we borrow ‘associated’ metadata to indicate the relationship for ".ref". There are three types of association need be addressed:
(1) variable to init/term functions (required to add functions to _cdtors array if the variable is included in linker output)
(2) dtor function to term function (required to correctly handle atexit/unatexit registration/unregistration)
(3) program code csect to internal variable classified as Data or ThreadData

This is the first half change to support -bcdtors:csect linker option, it adds ".ref" in case (3), and emit ".ref" in assembly output. The other two items will be handled in clang.

[shchenz]

llvm/include/llvm/MC/MCStreamer.h

625 ↗

(On Diff #417140)

Can we only use one parameter const MCSymbol *Symbol for this function? And update other callers accordingly. These two parameters are kind of duplicate.

llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp

2335

However .ref does not support BS mapping class.

This seems weird. Should be "csect with BS storage mapping class can not contain .ref"?

llvm/lib/Target/PowerPC/PPCAsmPrinter.cpp

2381

can we change the GVkind here for static globals with associated metadata? We use BSS kind for normal static globals, maybe we can change it to RW, so that we may don't need to change MCSectionXCOFF::printSwitchToSection and TargetLoweringObjectFileXCOFF::SelectSectionForGlobal?

2407

nit: !GV->getMetadata(LLVMContext::MD_associated) seems more common

[shchenz]

llvm/lib/MC/MCAsmStreamer.cpp

942

Maybe we can add an assert here that we are not adding .ref in a section with BS/UC storage mapping class.

[tingwang]

Update based on Chen Zheng's comments:
(1) Corrected emitXCOFFRefDirective argument, the original one should be fine.
(2) Add assert check to guard against BS and UC mapping class.
(3) Updated code comment and code style.

llvm/lib/Target/PowerPC/PPCAsmPrinter.cpp

2381

My initial thought was the same. After looking into TargetLoweringObjectFile::getKindForGlobal(), I didn't find any point for platform specific logic to change the decision, and thus I turned into our platform dependent code to make changes here. (If there is already one hook, I will be happy to use that. However at this moment, I don't see any for this purpose.)

[shchenz]

llvm/lib/MC/MCAsmStreamer.cpp

937

can we use getCurrentSectionOnly()?

llvm/lib/Target/PowerPC/PPCAsmPrinter.cpp

2381

Is it possible to custom the GVKind in current function? This function is AIX specific?

[tingwang]

llvm/lib/Target/PowerPC/PPCAsmPrinter.cpp

2381

If I'm understanding correct, this is customize in the middle (my original approach is customize in the end). The trace log shows that getKindForGlobal gets called from different places, for example: GlobalMerge.cpp:::GlobalMerge::doInitialization,
PPCAsmPrinter.cpp:::PPCAIXAsmPrinter::doInitialization, TargetLoweringObjectFile.cpp:::TargetLoweringObjectFile::SectionForGlobal, TargetLoweringObjectFileImpl.cpp:::TargetLoweringObjectFileXCOFF::getTargetSymbol. Customizing in the middle means we have to cover all above code path for correct behavior, and we also have to take care of any more upcoming code changes regarding this API usage. Probably the original approach is simpler.

[tingwang]

Add assert to guard MCSection

[tingwang]

During test some case revealed another issue regarding BSSLocal: there maybe some function associated with variable which has storage class BSSLocal (zero initialized). Unfortunately the variable is the association target, and there is no associated metadata attached to the variable. I have to work out a solution for this, incorporating Zheng's comments here. I will update code once the solution is formed.

[tingwang]

Use opaque-pointer in testcase aix-xcoff-ref.ll

[tingwang]

Rebase && drop TLS related .ref for now

[tingwang]

Rebase.

[stephenpeckham]

I'm not familiar with a lot of this code, but I don't see any obvious errors.

[tingwang]

In D122198#4185848, @stephenpeckham wrote:

I'm not familiar with a lot of this code, but I don't see any obvious errors.

Thank you for help!

[tingwang]

Rebase.

llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp

MCSection *TargetLoweringObjectFileXCOFF::SelectSectionForGlobal(
 
   // Common symbols go into a csect with matching name which will get mapped
   // into the .bss section.
   // Zero-initialized local TLS symbols go into a csect with matching name which
   // will get mapped into the .tbss section.
   if (Kind.isBSSLocal() || GO->hasCommonLinkage() || Kind.isThreadBSSLocal()) {
     SmallString<128> Name;
     getNameWithPrefix(Name, GO, TM);
-    XCOFF::StorageMappingClass SMC = Kind.isBSSLocal() ? XCOFF::XMC_BS
-                                     : Kind.isCommon() ? XCOFF::XMC_RW
-                                                       : XCOFF::XMC_UL;
+    bool EmitREF = false;
+    // Static global variable has associated initializer emits .ref to indicate
+    // the relationship. However csect with BS storage mapping class cannot
+    // contain .ref.
+    if (Kind.isBSSLocal() && GO->getMetadata(LLVMContext::MD_associated))
+      EmitREF = true;
+
+    XCOFF::StorageMappingClass SMC =
+        Kind.isBSSLocal() ? (EmitREF ? XCOFF::XMC_RW : XCOFF::XMC_BS)
     return getContext().getXCOFFSection(
-        Name, Kind, XCOFF::CsectProperties(SMC, XCOFF::XTY_CM));
+        Name, Kind,
+        XCOFF::CsectProperties(SMC, EmitREF ? XCOFF::XTY_SD : XCOFF::XTY_CM));
   }
 
   if (Kind.isMergeableCString()) {
     Align Alignment = GO->getParent()->getDataLayout().getPreferredAlign(
         cast<GlobalVariable>(GO));
 
     unsigned EntrySize = getEntrySizeForKind(Kind);
     std::string SizeSpec = ".rodata.str" + utostr(EntrySize) + ".";

llvm/lib/MC/MCAsmStreamer.cpp

if (C == DQ)
       OS << DQ;
     OS << C;
   }
   OS << DQ;
   EmitEOL();
 }
 
 void MCAsmStreamer::emitXCOFFRefDirective(StringRef Name) {
+  assert(
+      !(Name.endswith_insensitive("[bs]") ||
+        Name.endswith_insensitive("[uc]")) &&
+      ".ref operand cannot be a name with storage mapping class of BS or UC");
   OS << "\t.ref " << Name;
   EmitEOL();
 }
 
 void MCAsmStreamer::emitELFSize(MCSymbol *Symbol, const MCExpr *Value) {
   assert(MAI->hasDotTypeDotSizeDirective());
   OS << "\t.size\t";
   Symbol->print(OS, MAI);
   OS << ", ";
   Value->print(OS, MAI);
   EmitEOL();

llvm/lib/MC/MCSectionXCOFF.cpp

void MCSectionXCOFF::printSwitchToSection(const MCAsmInfo &MAI, const Triple &T,
   // XCOFF debug sections.
   if (getKind().isMetadata() && isDwarfSect()) {
     OS << "\n\t.dwsect "
        << format("0x%" PRIx32, getDwarfSubtypeFlags().getValue()) << '\n';
     OS << MAI.getPrivateLabelPrefix() << getName() << ':' << '\n';
     return;
   }
 
+  // Static global variable with initializer should map to RW
+  if (getKind().isBSSLocal() && isCsect() &&
+      getMappingClass() == XCOFF::XMC_RW && getCSectType() == XCOFF::XTY_SD) {
+    printCsectDirective(OS);
+    return;
+  }
+
   report_fatal_error("Printing for this SectionKind is unimplemented.");
 }
 
 bool MCSectionXCOFF::useCodeAlign() const { return getKind().isText(); }
 
 bool MCSectionXCOFF::isVirtualSection() const {
   // DWARF sections are always not virtual.
   if (isDwarfSect())
     return false;
   assert(isCsect() &&
          "Handling for isVirtualSection not implemented for this section!");
   return XCOFF::XTY_CM == CsectProp->Type;
 }

llvm/lib/Target/PowerPC/PPCAsmPrinter.cpp

private:
 
   uint16_t getNumberOfVRSaved();
   void emitTracebackTable();
 
   SmallVector<const GlobalVariable *, 8> TOCDataGlobalVars;
 
   void emitGlobalVariableHelper(const GlobalVariable *);
 
+  void emitAssociatedRefs(const GlobalObject &);
+
 public:
   PPCAIXAsmPrinter(TargetMachine &TM, std::unique_ptr<MCStreamer> Streamer)
       : PPCAsmPrinter(TM, std::move(Streamer)) {
     if (MAI->isLittleEndian())
       report_fatal_error(
           "cannot create AIX PPC Assembly Printer for a little-endian target");
   }
 
void PPCAIXAsmPrinter::emitGlobalVariableHelper(const GlobalVariable *GV) {
 
   MCSymbolXCOFF *GVSym = cast<MCSymbolXCOFF>(getSymbol(GV));
 
   if (GV->isDeclarationForLinker()) {
     emitLinkage(GV, GVSym);
     return;
   }
 
   SectionKind GVKind = getObjFileLowering().getKindForGlobal(GV, TM);
   if (!GVKind.isGlobalWriteableData() && !GVKind.isReadOnly() &&
       !GVKind.isThreadLocal()) // Checks for both ThreadData and ThreadBSS.
     report_fatal_error("Encountered a global variable kind that is "
                        "not supported yet.");
 
   // Print GV in verbose mode
   if (isVerbose()) {
     if (GV->hasInitializer()) {
       GV->printAsOperand(OutStreamer->GetCommentOS(),
                          /*PrintType=*/false, GV->getParent());
       OutStreamer->GetCommentOS() << '\n';
     }
   }
 
   MCSectionXCOFF *Csect = cast<MCSectionXCOFF>(
       getObjFileLowering().SectionForGlobal(GV, GVKind, TM));
 
   // Switch to the containing csect.
   OutStreamer->SwitchSection(Csect);
 
   const DataLayout &DL = GV->getParent()->getDataLayout();
 
   // Handle common and zero-initialized local symbols.
-  if (GV->hasCommonLinkage() || GVKind.isBSSLocal() ||
+  if (GV->hasCommonLinkage() ||
+      (GVKind.isBSSLocal() && !GV->getMetadata(LLVMContext::MD_associated)) ||
       GVKind.isThreadBSSLocal()) {
     Align Alignment = GV->getAlign().getValueOr(DL.getPreferredAlign(GV));
     uint64_t Size = DL.getTypeAllocSize(GV->getValueType());
     GVSym->setStorageClass(
         TargetLoweringObjectFileXCOFF::getStorageClassForGlobal(GV));
 
     if (GVKind.isBSSLocal() || GVKind.isThreadBSSLocal())
       OutStreamer->emitXCOFFLocalCommonSymbol(
           OutContext.getOrCreateSymbol(GVSym->getSymbolTableName()), Size,
           GVSym, Alignment.value());
     else
       OutStreamer->emitCommonSymbol(GVSym, Size, Alignment.value());
     return;
   }
 
   MCSymbol *EmittedInitSym = GVSym;
+  emitAssociatedRefs(*GV);
   emitLinkage(GV, EmittedInitSym);
   emitAlignment(getGVAlignment(GV, DL), GV);
 
   // When -fdata-sections is enabled, every GlobalVariable will
   // be put into its own csect; therefore, label is not necessary here.
   if (!TM.getDataSections() || GV->hasSection()) {
     OutStreamer->emitLabel(EmittedInitSym);
   }
 
   // Emit aliasing label for global variable.
   llvm::for_each(GOAliasMap[GV], [this](const GlobalAlias *Alias) {
     OutStreamer->emitLabel(getSymbol(Alias));
   });
 
   emitGlobalConstant(GV->getParent()->getDataLayout(), GV->getInitializer());
 }
 
+void PPCAIXAsmPrinter::emitAssociatedRefs(const GlobalObject &GO) {
+  MDNode *MD = GO.getMetadata(LLVMContext::MD_associated);
+  if (!MD)
+    return;
+
+  for (unsigned i = 0, e = MD->getNumOperands(); i != e; ++i) {
+    const MDOperand &Op = MD->getOperand(i);
+    if (!Op.get())
+      report_fatal_error("MD_associated operand not found");
+
+    const auto *VM = dyn_cast<ValueAsMetadata>(Op);
+    if (!VM)
+      report_fatal_error("MD_associated operand is not ValueAsMetadata");
+
+    if (const auto *OtherGV = dyn_cast<GlobalValue>(VM->getValue()))
+      OutStreamer->emitXCOFFRefDirective(getSymbol(OtherGV)->getName());
+  }
+}
+
 void PPCAIXAsmPrinter::emitFunctionDescriptor() {
   const DataLayout &DL = getDataLayout();
   const unsigned PointerSize = DL.getPointerSizeInBits() == 64 ? 8 : 4;
 
   MCSectionSubPair Current = OutStreamer->getCurrentSection();
   // Emit function descriptor.
   OutStreamer->SwitchSection(
       cast<MCSymbolXCOFF>(CurrentFnDescSym)->getRepresentedCsect());
 
   // Emit aliasing label for function descriptor csect.
   llvm::for_each(GOAliasMap[&MF->getFunction()],
                  [this](const GlobalAlias *Alias) {
                    OutStreamer->emitLabel(getSymbol(Alias));
                  });
 
+  // Emit reference to function or static variables
+  emitAssociatedRefs(MF->getFunction());
   // Emit function entry point address.
   OutStreamer->emitValue(MCSymbolRefExpr::create(CurrentFnSym, OutContext),
                          PointerSize);
   // Emit TOC base address.
   const MCSymbol *TOCBaseSym =
       cast<MCSectionXCOFF>(getObjFileLowering().getTOCBaseSection())
           ->getQualNameSymbol();
   OutStreamer->emitValue(MCSymbolRefExpr::create(TOCBaseSym, OutContext),

llvm/test/CodeGen/PowerPC/aix-xcoff-ref.ll

+; RUN: llc -mtriple powerpc-ibm-aix-xcoff --filetype=asm < %s | FileCheck %s
+
+@a = internal global i32 0, align 4, !associated !0
+@b = internal global i32 0, align 4, !associated !0
+@c = internal global i32 0, align 4
+
+define i32 @foo() !associated !1 {
+entry:
+  %0 = load i32, i32* @a, align 4
+  ret i32 %0
+}
+
+define void @bar() !associated !2 {
+entry:
+  ret void
+}
+
+define void @baz() !associated !3 {
+entry:
+  ret void
+}
+
+!0 = !{i32 ()* @foo, void ()* @bar}
+!1 = !{i32* @b}
+!2 = !{i32* @a, i32* @b}
+!3 = !{void ()* @bar}
+
+; CHECK:      .csect foo[DS],2
+; CHECK-NEXT: .ref    b[RW]                           # @foo
+
+; CHECK:      .csect bar[DS],2
+; CHECK-NEXT: .ref    a[RW]                           # @bar
+; CHECK-NEXT: .ref    b[RW]
+
+; CHECK:      .csect baz[DS],2
+; CHECK-NEXT: .ref    bar[DS]                         # @baz
+
+; CHECK:      .csect a[RW],2
+; CHECK-NEXT: .ref    foo[DS]                         # @a
+; CHECK-NEXT: .ref    bar[DS]
+
+; CHECK:      .csect b[RW],2
+; CHECK-NEXT: .ref    foo[DS]                         # @b
+; CHECK-NEXT: .ref    bar[DS]