[ELF][MIPS] Prevent substitution of _gp_disp symbol

[atanasyan]

Summary

On MIPS O32 ABI, _gp_disp is a magic symbol designates offset between start of function and gp pointer into GOT. To make seal with such symbol we add new method addIgnoredStrong(). It adds ignored symbol with global binding to prevent the symbol substitution. The addIgnored call is not enough here because this call adds a weak symbol which might be substituted by symbol from shared library.

[ruiu]

Is this behavior defined in the MIPS ABI, or is this just conventional?

Why addIgnored() didn't work? I think addIgnored() adds an absolute symbol to the symbol table, so it doesn't seem much different from this code.

test/ELF/mips-gp-disp.s

5 ↗

(On Diff #44548)

Can you create this so file in this file from an assembly file, instead of checking in a binary file?

[atanasyan]

In D16084#324046, @ruiu wrote:

Is this behavior defined in the MIPS ABI, or is this just conventional?

Why addIgnored() didn't work? I think addIgnored() adds an absolute symbol to the symbol table, so it doesn't seem much different from this code.

MIPS ABI says nothing about emitting _gp_disp symbol into the symbol table. But almost all legacy shared library has this symbol in their symbol tables. I guess by historical reasons. If I use addIgnored call the IgnoreUndef symbol has STB_WEAK binding. So corresponding MipsGpDisp SymbolBody it can be substituted by _gp_disp symbol defined in such shared library.

When I add MipsGpDisp with STB_GLOBAL binding the symbol gets priority over symbols from shared libraries.

test/ELF/mips-gp-disp.s

5 ↗

(On Diff #44548)

I can't. This so file has _gp_disp in the symbol table and emulates a legacy shared library. In fact we do not need to have _gp_disp in a symbol table and LLD cannot produce such shared library itself.

[ruiu]

Can you then add addIgnoredStrong() and IgnoredUndefStrong? The new ignored symbol is not actually specific to Mips or Mips _gp_disp symbol.

test/ELF/mips-gp-disp.s

5 ↗

(On Diff #44548)

Ah, interesting, but there's still some way to create such .so file in an unusual way. For example, you can create an .so with a dummy symbol _XX_XXXX and replace it with _gp_disp using sed.

[atanasyan]

• Add new general addIgnoredStrong() method
• Do not use binary files in the test

[ruiu]

LGTM

ELF/Driver.cpp

308 ↗

(On Diff #44565)

Can you add a comment on why you are using "Strong" as you described before?

test/ELF/Inputs/mips-gp-disp-def.s

1 ↗

(On Diff #44565)

Please add a comment why we are doing this (that we can't create a shared object file containing the special symbol.)

[atanasyan]

In D16084#324213, @ruiu wrote:

LGTM

By the way, do we still need to version of addIgnored method. What do you think, is it possible to always use the strong version?

[grimar]

Test fails under windows + msvs 2015:

21> FAIL: lld :: ELF/mips-gp-disp.s (213 of 1029)
21> **** TEST 'lld :: ELF/mips-gp-disp.s' FAILED ****
21> Script:
21> --
21> C:/access_softek/c_make_build_dir/Debug/bin\llvm-mc.EXE -filetype=obj -triple=mips-unknown-linux C:\access_softek\llvm\tools\lld\test\ELF/Inputs/mips-gp-disp-def.s -o C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext.o
21> C:/access_softek/c_make_build_dir/Debug/bin\ld.lld.EXE -shared -o C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext-int.so C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext.o
21> sed -e 's/XXXXXXXX/_gp_disp/g' C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext-int.so > C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext.so
21> C:/access_softek/c_make_build_dir/Debug/bin\llvm-mc.EXE -filetype=obj -triple=mips-unknown-linux C:\access_softek\llvm\tools\lld\test\ELF\mips-gp-disp.s -o C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp.o
21> C:/access_softek/c_make_build_dir/Debug/bin\ld.lld.EXE -shared -o C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp.so C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp.o C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext.so
21> C:/access_softek/c_make_build_dir/Debug/bin\llvm-readobj.EXE -symbols C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp.so | C:/access_softek/c_make_build_dir/Debug/bin\FileCheck.EXE -check-prefix=INT-SO C:\access_softek\llvm\tools\lld\test\ELF\mips-gp-disp.s
21> C:/access_softek/c_make_build_dir/Debug/bin\llvm-readobj.EXE -symbols C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext.so | C:/access_softek/c_make_build_dir/Debug/bin\FileCheck.EXE -check-prefix=EXT-SO C:\access_softek\llvm\tools\lld\test\ELF\mips-gp-disp.s
21> C:/access_softek/c_make_build_dir/Debug/bin\llvm-objdump.EXE -d -t C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp.so | C:/access_softek/c_make_build_dir/Debug/bin\FileCheck.EXE -check-prefix=DIS C:\access_softek\llvm\tools\lld\test\ELF\mips-gp-disp.s
21> --
21> Exit Code: 1
21>
21> Command Output (stdout):
21> --
21> Command 0: "C:/access_softek/c_make_build_dir/Debug/bin\llvm-mc.EXE" "-filetype=obj" "-triple=mips-unknown-linux" "C:\access_softek\llvm\tools\lld\test\ELF/Inputs/mips-gp-disp-def.s" "-o" "C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext.o"
21> Command 0 Result: 0
21> Command 0 Output:
21>
21>
21> Command 0 Stderr:
21>
21>
21> Command 1: "C:/access_softek/c_make_build_dir/Debug/bin\ld.lld.EXE" "-shared" "-o" "C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext-int.so" "C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext.o"
21> Command 1 Result: 0
21> Command 1 Output:
21>
21>
21> Command 1 Stderr:
21>
21>
21> Command 2: "sed" "-e" "s/XXXXXXXX/_gp_disp/g" "C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext-int.so"
21> Command 2 Result: 0
21> Command 2 Output:
21> None
21>
21> Command 2 Stderr:
21>
21>
21> Command 3: "C:/access_softek/c_make_build_dir/Debug/bin\llvm-mc.EXE" "-filetype=obj" "-triple=mips-unknown-linux" "C:\access_softek\llvm\tools\lld\test\ELF\mips-gp-disp.s" "-o" "C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp.o"
21> Command 3 Result: 0
21> Command 3 Output:
21>
21>
21> Command 3 Stderr:
21>
21>
21> Command 4: "C:/access_softek/c_make_build_dir/Debug/bin\ld.lld.EXE" "-shared" "-o" "C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp.so" "C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp.o" "C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext.so"
21> Command 4 Result: 1
21> Command 4 Output:
21>
21>
21> Command 4 Stderr:
21> Invalid data was encountered while parsing the file

[atanasyan]

In D16084#324527, @grimar wrote:

Test fails under windows + msvs 2015:

21> Command 3: "C:/access_softek/c_make_build_dir/Debug/bin\llvm-mc.EXE" "-filetype=obj" "-triple=mips-unknown-linux" "C:\access_softek\llvm\tools\lld\test\ELF\mips-gp-disp.s" "-o" "C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp.o"
21> Command 3 Result: 0
21> Command 3 Output:

Could you please send "C:\access_softek\c_make_build_dir\tools\lld\test\ELF\Output\mips-gp-disp.s.tmp-ext-int.so"? I guess sed corrupts the file for some reason.

[grimar]

• mips-gp-disp.s.tmp-ext-int.so128 KBDownload

[atanasyan]

The problem with failed test case on Windows fixed at rL257471.

lld/trunk/ELF/Driver.cpp

if (!Config->Entry.empty()) {
     // Set either EntryAddr (if S is a number) or EntrySym (otherwise).
     StringRef S = Config->Entry;
     if (S.getAsInteger(0, Config->EntryAddr))
       Config->EntrySym = Symtab.addUndefined(S);
   }
 
   if (Config->EMachine == EM_MIPS) {
     // On MIPS O32 ABI, _gp_disp is a magic symbol designates offset between
-    // start of function and gp pointer into GOT.
-    Config->MipsGpDisp = Symtab.addIgnored("_gp_disp");
+    // start of function and gp pointer into GOT. Use 'strong' variant of
+    // the addIgnored to prevent '_gp_disp' substitution.
+    Config->MipsGpDisp = Symtab.addIgnoredStrong("_gp_disp");
 
     // 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.
     // See "Global Data Symbols" in Chapter 6 in the following document:
     // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
     Symtab.addAbsolute("_gp", ElfSym<ELFT>::MipsGp);
   }
 

lld/trunk/ELF/SymbolTable.h

public:
   }
 
   SymbolBody *addUndefined(StringRef Name);
   SymbolBody *addUndefinedOpt(StringRef Name);
   SymbolBody *addAbsolute(StringRef Name, Elf_Sym &ESym);
   SymbolBody *addSynthetic(StringRef Name, OutputSectionBase<ELFT> &Section,
                            uintX_t Value);
   SymbolBody *addIgnored(StringRef Name);
+  SymbolBody *addIgnoredStrong(StringRef Name);
 
   void scanShlibUndefined();
   SymbolBody *find(StringRef Name);
   void wrap(StringRef Name);
   ELFFileBase<ELFT> *findFile(SymbolBody *B);
 
 private:
   Symbol *insert(SymbolBody *New);

lld/trunk/ELF/SymbolTable.cpp

 // Add Name as an "ignored" symbol. An ignored symbol is a regular
 // linker-synthesized defined symbol, but it is not recorded to the output
 // file's symbol table. Such symbols are useful for some linker-defined symbols.
 template <class ELFT>
 SymbolBody *SymbolTable<ELFT>::addIgnored(StringRef Name) {
   return addAbsolute(Name, ElfSym<ELFT>::IgnoreUndef);
 }
 
+// The 'strong' variant of the addIgnored. Adds symbol which has a global
+// binding and cannot be substituted.
+template <class ELFT>
+SymbolBody *SymbolTable<ELFT>::addIgnoredStrong(StringRef Name) {
+  return addAbsolute(Name, ElfSym<ELFT>::IgnoreUndefStrong);
+}
+
 // Rename SYM as __wrap_SYM. The original symbol is preserved as __real_SYM.
 // Used to implement --wrap.
 template <class ELFT> void SymbolTable<ELFT>::wrap(StringRef Name) {
   if (Symtab.count(Name) == 0)
     return;
   StringSaver Saver(Alloc);
   Symbol *Sym = addUndefined(Name)->getSymbol();
   Symbol *Real = addUndefined(Saver.save("__real_" + Name))->getSymbol();

lld/trunk/ELF/Symbols.h

 
 // Some linker-generated symbols need to be created as
 // DefinedRegular symbols, so they need Elf_Sym symbols.
 // Here we allocate such Elf_Sym symbols statically.
 template <class ELFT> struct ElfSym {
   typedef typename llvm::object::ELFFile<ELFT>::Elf_Sym Elf_Sym;
 
   // Used to represent an undefined symbol which we don't want
-  // to add to the output file's symbol table.
+  // to add to the output file's symbol table. The `IgnoreUndef`
+  // has weak binding and can be substituted. The `InoreUndefStrong`
+  // has global binding and gets priority over symbols from shared libs.
   static Elf_Sym IgnoreUndef;
+  static Elf_Sym IgnoreUndefStrong;
 
   // The content for _end and end symbols.
   static Elf_Sym End;
 
   // The content for _gp symbol for MIPS target.
   static Elf_Sym MipsGp;
 
   // __rel_iplt_start/__rel_iplt_end for signaling
   // where R_[*]_IRELATIVE relocations do live.
   static Elf_Sym RelaIpltStart;
   static Elf_Sym RelaIpltEnd;
 };
 
 template <class ELFT> typename ElfSym<ELFT>::Elf_Sym ElfSym<ELFT>::IgnoreUndef;
+template <class ELFT>
+typename ElfSym<ELFT>::Elf_Sym ElfSym<ELFT>::IgnoreUndefStrong;
 template <class ELFT> typename ElfSym<ELFT>::Elf_Sym ElfSym<ELFT>::End;
 template <class ELFT> typename ElfSym<ELFT>::Elf_Sym ElfSym<ELFT>::MipsGp;
 template <class ELFT>
 typename ElfSym<ELFT>::Elf_Sym ElfSym<ELFT>::RelaIpltStart;
 template <class ELFT> typename ElfSym<ELFT>::Elf_Sym ElfSym<ELFT>::RelaIpltEnd;
 
 } // namespace elf2
 } // namespace lld
 
 #endif

lld/trunk/ELF/Symbols.cpp

if (MBRef.getBuffer().empty())
     return std::unique_ptr<InputFile>(nullptr);
   return createObjectFile(MBRef);
 }
 
 template <class ELFT> static void doInitSymbols() {
   ElfSym<ELFT>::End.setBinding(STB_GLOBAL);
   ElfSym<ELFT>::IgnoreUndef.setBinding(STB_WEAK);
   ElfSym<ELFT>::IgnoreUndef.setVisibility(STV_HIDDEN);
+  ElfSym<ELFT>::IgnoreUndefStrong.setBinding(STB_GLOBAL);
+  ElfSym<ELFT>::IgnoreUndefStrong.setVisibility(STV_HIDDEN);
 }
 
 void elf2::initSymbols() {
   doInitSymbols<ELF32LE>();
   doInitSymbols<ELF32BE>();
   doInitSymbols<ELF64LE>();
   doInitSymbols<ELF64BE>();
 }

lld/trunk/ELF/Writer.cpp

 }
 
 template <class ELFT> static bool includeInSymtab(const SymbolBody &B) {
   if (!B.isUsedInRegularObj())
     return false;
 
   // Don't include synthetic symbols like __init_array_start in every output.
   if (auto *U = dyn_cast<DefinedRegular<ELFT>>(&B))
-    if (&U->Sym == &ElfSym<ELFT>::IgnoreUndef)
+    if (&U->Sym == &ElfSym<ELFT>::IgnoreUndef ||
+        &U->Sym == &ElfSym<ELFT>::IgnoreUndefStrong)
       return false;
 
   return true;
 }
 
 static bool includeInDynamicSymtab(const SymbolBody &B) {
   uint8_t V = B.getVisibility();
   if (V != STV_DEFAULT && V != STV_PROTECTED)

lld/trunk/test/ELF/Inputs/mips-gp-disp-def.s

+# We cannot create a shared library with defined _gp_disp symbol
+# so we use a workaround - create a library with XXXXXXXX symbols
+# and use 'sed' to replace it by _gp_disp right in the binary file.
+  .data
+  .globl XXXXXXXX
+  .space 16
+XXXXXXXX:
+  .space 4

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

+# Check that even if _gp_disp symbol is defined in the shared library
+# we use our own value.
+
+# RUN: llvm-mc -filetype=obj -triple=mips-unknown-linux \
+# RUN:         %S/Inputs/mips-gp-disp-def.s -o %t-ext.o
+# RUN: ld.lld -shared -o %t-ext-int.so %t-ext.o
+# RUN: sed -e 's/XXXXXXXX/_gp_disp/g' %t-ext-int.so > %t-ext.so
+# RUN: llvm-mc -filetype=obj -triple=mips-unknown-linux %s -o %t.o
+# RUN: ld.lld -shared -o %t.so %t.o %t-ext.so
+# RUN: llvm-readobj -symbols %t.so | FileCheck -check-prefix=INT-SO %s
+# RUN: llvm-readobj -symbols %t-ext.so | FileCheck -check-prefix=EXT-SO %s
+# RUN: llvm-objdump -d -t %t.so | FileCheck -check-prefix=DIS %s
+
+# REQUIRES: mips
+
+# INT-SO-NOT:  Name: _gp_disp
+
+# EXT-SO:      Name: _gp_disp
+# EXT-SO-NEXT: Value: 0x20010
+
+# DIS:      Disassembly of section .text:
+# DIS-NEXT: __start:
+# DIS-NEXT:    10000:  3c 08 00 01  lui   $8, 1
+# DIS-NEXT:    10004:  21 08 7f f0  addi  $8, $8, 32752
+#                                                 ^-- 0x37ff0 & 0xffff
+# DIS: 00027ff0  *ABS*  00000000 _gp
+
+  .text
+  .globl  __start
+__start:
+  lui    $t0,%hi(_gp_disp)
+  addi   $t0,$t0,%lo(_gp_disp)
+  lw     $v0,%call16(_foo)($gp)