[ELF] Resolve weak undefined TLS symbols when no phdr is available

[davide]

Summary

If we pass --gc-sections to lld and .tbss is not referenced, it gets reclaimed and we don't create a TLS program header.
R_TLS tries to access the program header -> lld crashes. Mimic what bfd/gold do in this case and resolve a weak undefined TLS symbol to the base of the TLS block, i.e. give it a value of zero.

[ruiu]

ELF/InputSection.cpp

249 ↗

(On Diff #72185)

We should never call getSymVA() on a lazy symbol, so isLazy() is always false. But if so, your code cannot change the behavior of this function. So something's not correct.

[davide]

I think we treat weak undefined specially. http://llvm.org/viewvc/llvm-project?view=revision&revision=261591

isLazy() returns true. At least it did when I stepped through inside a debugger.

[davide]

I don't think the invariant you point out is respected (we should never call getSymVA() on a Lazy symbol).
I was lazy, and to convince myself I put the following assertion in the code:

diff --git a/ELF/InputSection.cpp b/ELF/InputSection.cpp
index 3cc6ce5..b44d2d9 100644
--- a/ELF/InputSection.cpp
+++ b/ELF/InputSection.cpp
@@ -185,6 +185,7 @@ template <class ELFT>
 static typename ELFT::uint getSymVA(uint32_t Type, typename ELFT::uint A,
                                     typename ELFT::uint P,
                                     const SymbolBody &Body, RelExpr Expr) {
+  assert (!Body.isLazy() && "patatino");
   switch (Expr) {
   case R_HINT:
     llvm_unreachable("cannot relocate hint relocs");

and saw three tests failing:

********************
Testing Time: 17.63s
********************
Failing Tests (3):
    lld :: ELF/archive.s
    lld :: ELF/lto/undef-weak.ll
    lld :: ELF/tls-in-archive.s

Still, maybe this is not the best place to check.

[davide]

Update after Rafael's feedback.

[• rafael]

LGTM

lld/trunk/ELF/InputSection.cpp

static typename ELFT::uint getSymVA(uint32_t Type, typename ELFT::uint A,
   case R_GOTONLY_PC:
     return Out<ELFT>::Got->getVA() + A - P;
   case R_GOTONLY_PC_FROM_END:
     return Out<ELFT>::Got->getVA() + A - P + Out<ELFT>::Got->getSize();
   case R_RELAX_TLS_LD_TO_LE:
   case R_RELAX_TLS_IE_TO_LE:
   case R_RELAX_TLS_GD_TO_LE:
   case R_TLS:
+    // A weak undefined TLS symbol resolves to the base of the TLS
+    // block, i.e. gets a value of zero. If we pass --gc-sections to
+    // lld and .tbss is not referenced, it gets reclaimed and we don't
+    // create a TLS program header. Therefore, we resolve this
+    // statically to zero.
+    if (Body.isTls() && (Body.isLazy() || Body.isUndefined()) &&
+        Body.symbol()->isWeak())
+      return 0;
     if (Target->TcbSize)
       return Body.getVA<ELFT>(A) +
              alignTo(Target->TcbSize, Out<ELFT>::TlsPhdr->p_align);
     return Body.getVA<ELFT>(A) - Out<ELFT>::TlsPhdr->p_memsz;
   case R_RELAX_TLS_GD_TO_LE_NEG:
   case R_NEG_TLS:
     return Out<ELF32LE>::TlsPhdr->p_memsz - Body.getVA<ELFT>(A);
   case R_ABS:

lld/trunk/test/ELF/tls-weak-undef.s

+// REQUIRES: x86
+// RUN: llvm-mc -filetype=obj -triple=x86_64-pc-linux %s -o %t.o
+// RUN: ld.lld %t.o -o %t --gc-sections
+
+// RUN: llvm-mc -filetype=obj -triple=x86_64-pc-linux \
+// RUN:   %p/Inputs/tls-in-archive.s -o %t1.o
+// RUN: llvm-ar cru %t.a %t1.o
+// RUN: ld.lld %t.o %t.a -o %t
+
+// Check that lld doesn't crash because we don't reference
+// the TLS phdr when it's not created.
+        .globl  _start
+_start:
+        movq    foo@gottpoff(%rip), %rax
+        .section        .tbss,"awT",@nobits
+        .weak   foo