[ELF][ARM] Merge handleARMTlsRelocation() into handleTlsRelocation()

[MaskRay]

Summary

ARM and RISC-V do not support TLS relaxations. However, for General
Dynamic and Local Dynamic models, if we are producing an executable and
the symbol is non-preemptable, we know it must be defined and the
R_ARM_TLS_DTPMOD32/R_RISCV_TLS_DTPMOD{32,64} dynamic relocation can be
omitted because it is always 1.

Merge handleARMTlsRelocation() into handleTlsRelocation(). This
requires to add logic to R_TLSGD_PC and R_TLSLD_PC. As a bonus, the
additional logic in R_TLSGD_PC code can be shared by the TLS support for
RISC-V (D63220).

[peter.smith]

I've got some stylistic suggestions; otherwise this looks to do the same thing as the previous code.

ELF/Relocations.cpp

180 ↗

(On Diff #204760)

I think CanRelax would be more appropriate. To me Relax implies the action of relaxation rather than whether we can take the action.

189 ↗

(On Diff #204760)

I think this reduces to !Config->Shared as this implies !Sym.IsPreemptible (see computeIsPreemptible).

189 ↗

(On Diff #204760)

I'd prefer not to use IsLocalExec to avoid confusing it with local exec model TLS, which I believe is just TPREL relative relocations that can be resolved statically.

In this context setting the module id to 1 is used in General and Local Dynamic models.

248 ↗

(On Diff #204760)

Is it possible to remove the following from Arch/ARM.cpp:

case R_ARM_TLS_DTPMOD32:
  write32le(Loc, 1);
  break;

As this used to be handled in handleARMTlsRelocation with:

AddTlsReloc(In.Got->getTlsIndexOff(), Target->TlsModuleIndexRel,
            NeedDynId ? nullptr : &Sym, NeedDynId);

[MaskRay]

Address peter.smith's comments

[MaskRay]

ELF/Relocations.cpp

189 ↗

(On Diff #204760)

bool IsLocalExec = !Sym.IsPreemptible && !Config->Shared; can be written as

bool IsLocalExec = Sym.isDefined() && !Config->Shared;

but it can't reduce to:

bool IsLocalExec = !Config->Shared;

The symbol may be a Shared, in that case the symbol isn't considered local.

IsLocalExec does seem confusing. I'll rename it to IsLocalInExecutable

248 ↗

(On Diff #204760)

Thanks! I noticed the R_RISCV_TLS_DTPMOD{32,64} can be omitted but somehow forget to delete it from ARM.cpp ...

[MaskRay]

Delete the word relocations from a comment.

[peter.smith]

Happy with the changes, thanks for the update.

[ruiu]

LGTM

lld/trunk/ELF/Arch/ARM.cpp

void ARM::relocateOne(uint8_t *Loc, RelType Type, uint64_t Val) const {
   case R_ARM_TLS_IE32:
   case R_ARM_TLS_LDM32:
   case R_ARM_TLS_LDO32:
   case R_ARM_TLS_LE32:
   case R_ARM_TLS_TPOFF32:
   case R_ARM_TLS_DTPOFF32:
     write32le(Loc, Val);
     break;
-  case R_ARM_TLS_DTPMOD32:
-    write32le(Loc, 1);
-    break;
   case R_ARM_PREL31:
     checkInt(Loc, Val, 31, Type);
     write32le(Loc, (read32le(Loc) & 0x80000000) | (Val & ~0x80000000));
     break;
   case R_ARM_CALL:
     // R_ARM_CALL is used for BL and BLX instructions, depending on the
     // value of bit 0 of Val, we must select a BL or BLX instruction
     if (Val & 1) {

lld/trunk/ELF/Relocations.cpp

static unsigned handleMipsTlsRelocation(RelType Type, Symbol &Sym,
   if (Expr == R_MIPS_TLSGD) {
     In.MipsGot->addDynTlsEntry(*C.File, Sym);
     C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
     return 1;
   }
   return 0;
 }
 
-// This function is similar to the `handleMipsTlsRelocation`. ARM also does not
-// support any relaxations for TLS relocations. ARM is logically similar to Mips
-// in how it handles TLS, but Mips uses its own custom GOT which handles some
-// of the cases that ARM uses GOT relocations for.
-//
-// We look for TLS global dynamic and local dynamic relocations, these may
-// require the generation of a pair of GOT entries that have associated
-// dynamic relocations. When the results of the dynamic relocations can be
-// resolved at static link time we do so. This is necessary for static linking
-// as there will be no dynamic loader to resolve them at load-time.
-//
-// The pair of GOT entries created are of the form
-// GOT[e0] Module Index (Used to find pointer to TLS block at run-time)
-// GOT[e1] Offset of symbol in TLS block
-static unsigned handleARMTlsRelocation(RelType Type, Symbol &Sym,
-                                       InputSectionBase &C, uint64_t Offset,
-                                       int64_t Addend, RelExpr Expr) {
-  // The Dynamic TLS Module Index Relocation for a symbol defined in an
-  // executable is always 1. If the target Symbol is not preemptible then
-  // we know the offset into the TLS block at static link time.
-  bool NeedDynId = Sym.IsPreemptible || Config->Shared;
-  bool NeedDynOff = Sym.IsPreemptible;
-
-  auto AddTlsReloc = [&](uint64_t Off, RelType Type, Symbol *Dest, bool Dyn) {
-    if (Dyn)
-      Main->RelaDyn->addReloc(Type, In.Got, Off, Dest);
-    else
-      In.Got->Relocations.push_back({R_ABS, Type, Off, 0, Dest});
-  };
-
-  // Local Dynamic is for access to module local TLS variables, while still
-  // being suitable for being dynamically loaded via dlopen.
-  // GOT[e0] is the module index, with a special value of 0 for the current
-  // module. GOT[e1] is unused. There only needs to be one module index entry.
-  if (Expr == R_TLSLD_PC && In.Got->addTlsIndex()) {
-    AddTlsReloc(In.Got->getTlsIndexOff(), Target->TlsModuleIndexRel,
-                NeedDynId ? nullptr : &Sym, NeedDynId);
-    C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
-    return 1;
-  }
-
-  // Global Dynamic is the most general purpose access model. When we know
-  // the module index and offset of symbol in TLS block we can fill these in
-  // using static GOT relocations.
-  if (Expr == R_TLSGD_PC) {
-    if (In.Got->addDynTlsEntry(Sym)) {
-      uint64_t Off = In.Got->getGlobalDynOffset(Sym);
-      AddTlsReloc(Off, Target->TlsModuleIndexRel, &Sym, NeedDynId);
-      AddTlsReloc(Off + Config->Wordsize, Target->TlsOffsetRel, &Sym,
-                  NeedDynOff);
-    }
-    C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
-    return 1;
-  }
-  return 0;
-}
-
+// Notes about General Dynamic and Local Dynamic TLS models below. They may
+// require the generation of a pair of GOT entries that have associated dynamic
+// relocations. The pair of GOT entries created are of the form GOT[e0] Module
+// Index (Used to find pointer to TLS block at run-time) GOT[e1] Offset of
+// symbol in TLS block.
+//
 // Returns the number of relocations processed.
 template <class ELFT>
 static unsigned
 handleTlsRelocation(RelType Type, Symbol &Sym, InputSectionBase &C,
                     typename ELFT::uint Offset, int64_t Addend, RelExpr Expr) {
   if (!Sym.isTls())
     return 0;
 
-  if (Config->EMachine == EM_ARM)
-    return handleARMTlsRelocation(Type, Sym, C, Offset, Addend, Expr);
   if (Config->EMachine == EM_MIPS)
     return handleMipsTlsRelocation(Type, Sym, C, Offset, Addend, Expr);
 
   if (oneof<R_AARCH64_TLSDESC_PAGE, R_TLSDESC, R_TLSDESC_CALL, R_TLSDESC_PC>(
           Expr) &&
       Config->Shared) {
     if (In.Got->addDynTlsEntry(Sym)) {
       uint64_t Off = In.Got->getGlobalDynOffset(Sym);
       Main->RelaDyn->addReloc(
           {Target->TlsDescRel, In.Got, Off, !Sym.IsPreemptible, &Sym, 0});
     }
     if (Expr != R_TLSDESC_CALL)
       C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
     return 1;
   }
 
+  bool CanRelax = Config->EMachine != EM_ARM;
+
+  // If we are producing an executable and the symbol is non-preemptable, it
+  // must be defined and the code sequence can be relaxed to use Local-Exec.
+  //
+  // ARM and RISC-V do not support any relaxations for TLS relocations, however,
+  // we can omit the DTPMOD dynamic relocations and resolve them at link time
+  // because them are always 1. This may be necessary for static linking as
+  // DTPMOD may not be expected at load time.
+  bool IsLocalInExecutable = !Sym.IsPreemptible && !Config->Shared;
+
+  // Local Dynamic is for access to module local TLS variables, while still
+  // being suitable for being dynamically loaded via dlopen. GOT[e0] is the
+  // module index, with a special value of 0 for the current module. GOT[e1] is
+  // unused. There only needs to be one module index entry.
   if (oneof<R_TLSLD_GOT, R_TLSLD_GOTPLT, R_TLSLD_PC, R_TLSLD_HINT>(
           Expr)) {
     // Local-Dynamic relocs can be relaxed to Local-Exec.
-    if (!Config->Shared) {
+    if (CanRelax && !Config->Shared) {
       C.Relocations.push_back(
           {Target->adjustRelaxExpr(Type, nullptr, R_RELAX_TLS_LD_TO_LE), Type,
            Offset, Addend, &Sym});
       return Target->getTlsGdRelaxSkip(Type);
     }
     if (Expr == R_TLSLD_HINT)
       return 1;
-    if (In.Got->addTlsIndex())
+    if (In.Got->addTlsIndex()) {
+      if (IsLocalInExecutable)
+        In.Got->Relocations.push_back(
+            {R_ADDEND, Target->SymbolicRel, In.Got->getTlsIndexOff(), 1, &Sym});
+      else
         Main->RelaDyn->addReloc(Target->TlsModuleIndexRel, In.Got,
                                 In.Got->getTlsIndexOff(), nullptr);
+    }
     C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
     return 1;
   }
 
   // Local-Dynamic relocs can be relaxed to Local-Exec.
   if (Expr == R_DTPREL && !Config->Shared) {
     C.Relocations.push_back(
         {Target->adjustRelaxExpr(Type, nullptr, R_RELAX_TLS_LD_TO_LE), Type,
if (!Sym.isInGot()) {
           {R_ABS, Target->TlsOffsetRel, Off, 0, &Sym});
     }
     C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
     return 1;
   }
 
   if (oneof<R_AARCH64_TLSDESC_PAGE, R_TLSDESC, R_TLSDESC_CALL, R_TLSDESC_PC,
             R_TLSGD_GOT, R_TLSGD_GOTPLT, R_TLSGD_PC>(Expr)) {
-    if (Config->Shared) {
+    if (!CanRelax || Config->Shared) {
       if (In.Got->addDynTlsEntry(Sym)) {
         uint64_t Off = In.Got->getGlobalDynOffset(Sym);
+
+        if (IsLocalInExecutable)
+          // Write one to the GOT slot.
+          In.Got->Relocations.push_back(
+              {R_ADDEND, Target->SymbolicRel, Off, 1, &Sym});
+        else
           Main->RelaDyn->addReloc(Target->TlsModuleIndexRel, In.Got, Off, &Sym);
 
         // If the symbol is preemptible we need the dynamic linker to write
         // the offset too.
         uint64_t OffsetOff = Off + Config->Wordsize;
         if (Sym.IsPreemptible)
           Main->RelaDyn->addReloc(Target->TlsOffsetRel, In.Got, OffsetOff,
                                   &Sym);
         else
if (oneof<R_AARCH64_TLSDESC_PAGE, R_TLSDESC, R_TLSDESC_CALL, R_TLSDESC_PC,
     }
     return Target->getTlsGdRelaxSkip(Type);
   }
 
   // Initial-Exec relocs can be relaxed to Local-Exec if the symbol is locally
   // defined.
   if (oneof<R_GOT, R_GOTPLT, R_GOT_PC, R_AARCH64_GOT_PAGE_PC, R_GOT_OFF,
             R_TLSIE_HINT>(Expr) &&
-      !Config->Shared && !Sym.IsPreemptible) {
+      CanRelax && IsLocalInExecutable) {
     C.Relocations.push_back({R_RELAX_TLS_IE_TO_LE, Type, Offset, Addend, &Sym});
     return 1;
   }
 
   if (Expr == R_TLSIE_HINT)
     return 1;
   return 0;
 }