[ELF][MIPS] Fix handling of _gp/_gp_disp/__gnu_local_gp symbols

[atanasyan]

Summary

Offset between beginning of a .got section and _gp symbols used in MIPS GOT relocations calculations. Usually the expression looks like VA + Offset - GP, where VA is the .got section address, Offset - offset of the GOT entry, GP - offset between .got and _gp. Also there two "magic" symbols _gp_disp and __gnu_local_gp which hold the offset mentioned above. These symbols might be referenced by MIPS relocations.

Now the linker always defines _gp symbol and uses hardcoded value for its initialization. So offset between .got and _gp is 0x7ff0. The _gp_disp and gnu_local_gp defined if required and initialized by 0x7ff0. In fact that is not correct because _gp symbol might be defined by a linker script and holds arbitrary value. In that case we need to use this value in relocation calculation and initialize _gp_disp and gnu_local_gp properly.

The patch fixes the problem and completes fixing the bug #30311.
https://llvm.org/bugs/show_bug.cgi?id=30311

[ruiu]

LGTM

ELF/SyntheticSections.h

125 ↗

(On Diff #79024)

nit: getGp for consistency.

ELF/Writer.cpp

1421 ↗

(On Diff #79024)

Compute this value only once.

lld/trunk/ELF/InputSection.cpp

static typename ELFT::uint getSymVA(uint32_t Type, typename ELFT::uint A,
   case R_RELAX_GOT_PC_NOPIC:
     return Body.getVA<ELFT>(A);
   case R_GOT_OFF:
     return Body.getGotOffset<ELFT>() + A;
   case R_MIPS_GOT_LOCAL_PAGE:
     // If relocation against MIPS local symbol requires GOT entry, this entry
     // should be initialized by 'page address'. This address is high 16-bits
     // of sum the symbol's value and the addend.
-    return In<ELFT>::MipsGot->getPageEntryOffset(Body.getVA<ELFT>(A));
+    return In<ELFT>::MipsGot->getVA() +
+           In<ELFT>::MipsGot->getPageEntryOffset(Body.getVA<ELFT>(A)) -
+           In<ELFT>::MipsGot->getGp();
   case R_MIPS_GOT_OFF:
   case R_MIPS_GOT_OFF32:
     // In case of MIPS if a GOT relocation has non-zero addend this addend
     // should be applied to the GOT entry content not to the GOT entry offset.
     // That is why we use separate expression type.
-    return In<ELFT>::MipsGot->getBodyEntryOffset(Body, A);
+    return In<ELFT>::MipsGot->getVA() +
+           In<ELFT>::MipsGot->getBodyEntryOffset(Body, A) -
+           In<ELFT>::MipsGot->getGp();
   case R_MIPS_GOTREL:
-    return Body.getVA<ELFT>(A) - In<ELFT>::MipsGot->getVA() - MipsGPOffset;
+    return Body.getVA<ELFT>(A) - In<ELFT>::MipsGot->getGp();
   case R_MIPS_TLSGD:
-    return In<ELFT>::MipsGot->getGlobalDynOffset(Body) +
-           In<ELFT>::MipsGot->getTlsOffset() - MipsGPOffset;
+    return In<ELFT>::MipsGot->getVA() + In<ELFT>::MipsGot->getTlsOffset() +
+           In<ELFT>::MipsGot->getGlobalDynOffset(Body) -
+           In<ELFT>::MipsGot->getGp();
   case R_MIPS_TLSLD:
-    return In<ELFT>::MipsGot->getTlsIndexOff() +
-           In<ELFT>::MipsGot->getTlsOffset() - MipsGPOffset;
+    return In<ELFT>::MipsGot->getVA() + In<ELFT>::MipsGot->getTlsOffset() +
+           In<ELFT>::MipsGot->getTlsIndexOff() - In<ELFT>::MipsGot->getGp();
   case R_PPC_OPD: {
     uint64_t SymVA = Body.getVA<ELFT>(A);
     // If we have an undefined weak symbol, we might get here with a symbol
     // address of zero. That could overflow, but the code must be unreachable,
     // so don't bother doing anything at all.
     if (!SymVA)
       return 0;
     if (Out<ELF64BE>::Opd) {

lld/trunk/ELF/Symbols.h

template <class ELFT> struct ElfSym {
   // The content for _edata and edata symbols.
   static DefinedRegular<ELFT> *Edata;
   static DefinedRegular<ELFT> *Edata2;
 
   // The content for _end and end symbols.
   static DefinedRegular<ELFT> *End;
   static DefinedRegular<ELFT> *End2;
 
-  // The content for _gp_disp symbol for MIPS target.
-  static SymbolBody *MipsGpDisp;
+  // The content for _gp_disp/__gnu_local_gp symbols for MIPS target.
+  static DefinedRegular<ELFT> *MipsGpDisp;
+  static DefinedRegular<ELFT> *MipsLocalGp;
+  static SymbolBody *MipsGp;
 };
 
 template <class ELFT> DefinedRegular<ELFT> *ElfSym<ELFT>::EhdrStart;
 template <class ELFT> DefinedRegular<ELFT> *ElfSym<ELFT>::Etext;
 template <class ELFT> DefinedRegular<ELFT> *ElfSym<ELFT>::Etext2;
 template <class ELFT> DefinedRegular<ELFT> *ElfSym<ELFT>::Edata;
 template <class ELFT> DefinedRegular<ELFT> *ElfSym<ELFT>::Edata2;
 template <class ELFT> DefinedRegular<ELFT> *ElfSym<ELFT>::End;
 template <class ELFT> DefinedRegular<ELFT> *ElfSym<ELFT>::End2;
-template <class ELFT> SymbolBody *ElfSym<ELFT>::MipsGpDisp;
+template <class ELFT> DefinedRegular<ELFT> *ElfSym<ELFT>::MipsGpDisp;
+template <class ELFT> DefinedRegular<ELFT> *ElfSym<ELFT>::MipsLocalGp;
+template <class ELFT> SymbolBody *ElfSym<ELFT>::MipsGp;
 
 // A real symbol object, SymbolBody, is usually stored within a Symbol. There's
 // always one Symbol for each symbol name. The resolver updates the SymbolBody
 // stored in the Body field of this object as it resolves symbols. Symbol also
 // holds computed properties of symbol names.
 struct Symbol {
   // Symbol binding. This is on the Symbol to track changes during resolution.
   // In particular:

lld/trunk/ELF/SyntheticSections.h

public:
   unsigned getLocalEntriesNum() const;
 
   // Returns offset of TLS part of the MIPS GOT table. This part goes
   // after 'local' and 'global' entries.
   uintX_t getTlsOffset() const;
 
   uint32_t getTlsIndexOff() const { return TlsIndexOff; }
 
+  unsigned getGp() const;
+
 private:
   // MIPS GOT consists of three parts: local, global and tls. Each part
   // contains different types of entries. Here is a layout of GOT:
   // - Header entries                |
   // - Page entries                  |   Local part
   // - Local entries (16-bit access) |
   // - Local entries (32-bit access) |
   // - Normal global entries         ||  Global part

lld/trunk/ELF/SyntheticSections.cpp

: SyntheticSection<ELFT>(SHF_ALLOC, SHT_MIPS_OPTIONS, 8, ".MIPS.options"),
       Reginfo(Reginfo) {}
 
 template <class ELFT> void MipsOptionsSection<ELFT>::writeTo(uint8_t *Buf) {
   auto *Options = reinterpret_cast<Elf_Mips_Options *>(Buf);
   Options->kind = ODK_REGINFO;
   Options->size = getSize();
 
   if (!Config->Relocatable)
-    Reginfo.ri_gp_value = In<ELFT>::MipsGot->getVA() + MipsGPOffset;
+    Reginfo.ri_gp_value = In<ELFT>::MipsGot->getGp();
   memcpy(Buf + sizeof(typename ELFT::uint), &Reginfo, sizeof(Reginfo));
 }
 
 template <class ELFT>
 MipsOptionsSection<ELFT> *MipsOptionsSection<ELFT>::create() {
   // N64 ABI only.
   if (!ELFT::Is64Bits)
     return nullptr;

 // MIPS .reginfo section.
 template <class ELFT>
 MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo Reginfo)
     : SyntheticSection<ELFT>(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo"),
       Reginfo(Reginfo) {}
 
 template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *Buf) {
   if (!Config->Relocatable)
-    Reginfo.ri_gp_value = In<ELFT>::MipsGot->getVA() + MipsGPOffset;
+    Reginfo.ri_gp_value = In<ELFT>::MipsGot->getGp();
   memcpy(Buf, &Reginfo, sizeof(Reginfo));
 }
 
 template <class ELFT>
 MipsReginfoSection<ELFT> *MipsReginfoSection<ELFT>::create() {
   // Section should be alive for O32 and N32 ABIs only.
   if (ELFT::Is64Bits)
     return nullptr;
MipsGotSection<ELFT>::getPageEntryOffset(uintX_t EntryValue) {
   // Initialize the entry by the %hi(EntryValue) expression
   // but without right-shifting.
   EntryValue = (EntryValue + 0x8000) & ~0xffff;
   // Take into account MIPS GOT header.
   // See comment in the MipsGotSection::writeTo.
   size_t NewIndex = PageIndexMap.size() + 2;
   auto P = PageIndexMap.insert(std::make_pair(EntryValue, NewIndex));
   assert(!P.second || PageIndexMap.size() <= PageEntriesNum);
-  return (uintX_t)P.first->second * sizeof(uintX_t) - MipsGPOffset;
+  return (uintX_t)P.first->second * sizeof(uintX_t);
 }
 
 template <class ELFT>
 typename MipsGotSection<ELFT>::uintX_t
 MipsGotSection<ELFT>::getBodyEntryOffset(const SymbolBody &B,
                                          uintX_t Addend) const {
   // Calculate offset of the GOT entries block: TLS, global, local.
   uintX_t GotBlockOff;
MipsGotSection<ELFT>::getBodyEntryOffset(const SymbolBody &B,
   uintX_t GotIndex;
   if (B.isInGot())
     GotIndex = B.GotIndex;
   else {
     auto It = EntryIndexMap.find({&B, Addend});
     assert(It != EntryIndexMap.end());
     GotIndex = It->second;
   }
-  return GotBlockOff + GotIndex * sizeof(uintX_t) - MipsGPOffset;
+  return GotBlockOff + GotIndex * sizeof(uintX_t);
 }
 
 template <class ELFT>
 typename MipsGotSection<ELFT>::uintX_t
 MipsGotSection<ELFT>::getTlsOffset() const {
   return (getLocalEntriesNum() + GlobalEntries.size()) * sizeof(uintX_t);
 }
 
for (const OutputSectionBase *OutSec : OutSections) {
     // Add 0x8000 to the section's size because the page address stored
     // in the GOT entry is calculated as (value + 0x8000) & ~0xffff.
     PageEntriesNum += (OutSec->Size + 0x8000 + 0xfffe) / 0xffff;
   }
   EntriesNum += getLocalEntriesNum() + GlobalEntries.size();
   Size = EntriesNum * sizeof(uintX_t);
 }
 
+template <class ELFT> unsigned MipsGotSection<ELFT>::getGp() const {
+  return ElfSym<ELFT>::MipsGp->template getVA<ELFT>(0);
+}
+
 template <class ELFT>
 static void writeUint(uint8_t *Buf, typename ELFT::uint Val) {
   typedef typename ELFT::uint uintX_t;
   write<uintX_t, ELFT::TargetEndianness, sizeof(uintX_t)>(Buf, Val);
 }
 
 template <class ELFT> void MipsGotSection<ELFT>::writeTo(uint8_t *Buf) {
   // Set the MSB of the second GOT slot. This is not required by any

lld/trunk/ELF/Target.h

public:
   virtual void relaxTlsGdToLe(uint8_t *Loc, uint32_t Type, uint64_t Val) const;
   virtual void relaxTlsIeToLe(uint8_t *Loc, uint32_t Type, uint64_t Val) const;
   virtual void relaxTlsLdToLe(uint8_t *Loc, uint32_t Type, uint64_t Val) const;
 };
 
 std::string toString(uint32_t RelType);
 uint64_t getPPC64TocBase();
 
-const unsigned MipsGPOffset = 0x7ff0;
-
 extern TargetInfo *Target;
 TargetInfo *createTarget();
 }
 }
 
 #endif

lld/trunk/ELF/Writer.cpp

 }
 
 // The linker is expected to define some symbols depending on
 // the linking result. This function defines such symbols.
 template <class ELFT> void Writer<ELFT>::addReservedSymbols() {
   if (Config->EMachine == EM_MIPS) {
     // Define _gp for MIPS. st_value of _gp symbol will be updated by Writer
     // so that it points to an absolute address which is relative to GOT.
+    // Default offset is 0x7ff0.
     // See "Global Data Symbols" in Chapter 6 in the following document:
     // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
-    addRegular("_gp", In<ELFT>::MipsGot, MipsGPOffset);
+    ElfSym<ELFT>::MipsGp = addRegular("_gp", In<ELFT>::MipsGot, 0x7ff0)->body();
 
     // On MIPS O32 ABI, _gp_disp is a magic symbol designates offset between
     // start of function and 'gp' pointer into GOT.
-    Symbol *Sym =
-        addOptionalRegular("_gp_disp", In<ELFT>::MipsGot, MipsGPOffset);
-    if (Sym)
-      ElfSym<ELFT>::MipsGpDisp = Sym->body();
+    if (Symbol *S = addOptionalRegular("_gp_disp", In<ELFT>::MipsGot, 0))
+      ElfSym<ELFT>::MipsGpDisp = cast<DefinedRegular<ELFT>>(S->body());
 
     // The __gnu_local_gp is a magic symbol equal to the current value of 'gp'
     // pointer. This symbol is used in the code generated by .cpload pseudo-op
     // in case of using -mno-shared option.
     // https://sourceware.org/ml/binutils/2004-12/msg00094.html
-    addOptionalRegular("__gnu_local_gp", In<ELFT>::MipsGot, MipsGPOffset);
+    if (Symbol *S = addOptionalRegular("__gnu_local_gp", In<ELFT>::MipsGot, 0))
+      ElfSym<ELFT>::MipsLocalGp = cast<DefinedRegular<ELFT>>(S->body());
   }
 
   // In the assembly for 32 bit x86 the _GLOBAL_OFFSET_TABLE_ symbol
   // is magical and is used to produce a R_386_GOTPC relocation.
   // The R_386_GOTPC relocation value doesn't actually depend on the
   // symbol value, so it could use an index of STN_UNDEF which, according
   // to the spec, means the symbol value is 0.
   // Unfortunately both gas and MC keep the _GLOBAL_OFFSET_TABLE_ symbol in
for (Phdr &P : Phdrs) {
     Set(ElfSym<ELFT>::End, ElfSym<ELFT>::End2, H.p_vaddr + H.p_memsz);
 
     uintX_t Val = H.p_vaddr + H.p_filesz;
     if (H.p_flags & PF_W)
       Set(ElfSym<ELFT>::Edata, ElfSym<ELFT>::Edata2, Val);
     else
       Set(ElfSym<ELFT>::Etext, ElfSym<ELFT>::Etext2, Val);
   }
+
+  // Setup MIPS _gp_disp/__gnu_local_gp symbols which should
+  // be equal to the _gp symbol's value.
+  if (Config->EMachine == EM_MIPS) {
+    uintX_t GpDisp = In<ELFT>::MipsGot->getGp() - In<ELFT>::MipsGot->getVA();
+    if (ElfSym<ELFT>::MipsGpDisp)
+      ElfSym<ELFT>::MipsGpDisp->Value = GpDisp;
+    if (ElfSym<ELFT>::MipsLocalGp)
+      ElfSym<ELFT>::MipsLocalGp->Value = GpDisp;
+  }
 }
 
 template <class ELFT> void Writer<ELFT>::writeHeader() {
   uint8_t *Buf = Buffer->getBufferStart();
   memcpy(Buf, "\177ELF", 4);
 
   // Write the ELF header.
   auto *EHdr = reinterpret_cast<Elf_Ehdr *>(Buf);

lld/trunk/test/ELF/mips-gp-ext.s

+# Check that the linker use a value of _gp symbol defined
+# in a linker script to calculate GOT relocations.
+
+# RUN: llvm-mc -filetype=obj -triple=mips-unknown-linux %s -o %t.o
+# RUN: echo "SECTIONS { \
+# RUN:          .text : { *(.text) } \
+# RUN:          _gp = . + 0x100; \
+# RUN:          .got  : { *(.got) } }" > %t.script
+# RUN: ld.lld -shared -o %t.so --script %t.script %t.o
+# RUN: llvm-objdump -s -t %t.so | FileCheck %s
+
+# REQUIRES: mips
+
+# CHECK:      Contents of section .text:
+# CHECK-NEXT:  0000 3c080000 2108010c 8f82ffe4
+#                   ^-- %hi(_gp_disp)
+#                            ^-- %lo(_gp_disp)
+#                                     ^-- 8 - (0x10c - 0xe8)
+#                                         G - (GP - .got)
+
+# CHECK:      Contents of section .reginfo:
+# CHECK-NEXT:  0028 10000104 00000000 00000000 00000000
+# CHECK-NEXT:  0038 00000000 0000010c
+#                            ^-- _gp
+
+# CHECK:      Contents of section .data:
+# CHECK-NEXT:  0100 fffffef4
+#                   ^-- 0-0x10c
+
+# CHECK: 00000000         .text           00000000 foo
+# CHECK: 0000010c         .got            00000000 .hidden _gp_disp
+# CHECK: 0000010c         .text           00000000 .hidden _gp
+
+  .text
+foo:
+  lui    $t0, %hi(_gp_disp)
+  addi   $t0, $t0, %lo(_gp_disp)
+  lw     $v0, %call16(bar)($gp)
+
+  .data
+  .gpword foo