[lld] [LLD] Tombstone LocalTU entry in .debug_names (PR #70701)

[Alexander Yermolovich via llvm-commits]

================
@@ -918,7 +917,8 @@ void InputSection::relocateNonAlloc(uint8_t *buf, ArrayRef<RelTy> rels) {
       continue;
 
     if (tombstone ||
-        (isDebug && (type == target.symbolicRel || expr == R_DTPREL))) {
+        ((isDebug && (type == target.symbolicRel || expr == R_DTPREL))) ||
----------------
ayermolo wrote:

OK I think I understand the history here....
My resources:
https://sourceware.org/pipermail/binutils/2020-May/111357.html
https://lists.llvm.org/pipermail/llvm-dev/2020-May/141885.html
https://reviews.llvm.org/D81784
https://reviews.llvm.org/D84825

TLDR on this whole thing. This code primarily deals with an issue of what to do with symbols in debug sections when text sections they reference get GC/ICF. Those addresses need to be tombsoned. The 0x0 doesn't work in .debug_ranges/.debug_loc because before DWARF5 0x0,0x0 is how those lists ended. The -1 doesn't work because it has special meaning. So first -2 was implemented, but that had issues also, so then it was changed to 1 for loc/ranges and 0 for "everything else".

Except since relocation type allowed (type == target.symbolicRel) is 64 bit "everything else" is .debug_addr section. Since in 32 bit dwarf only thing that is 64bit are addresses in .text section. At some point expr == R_DTPREL was added for TLS stuff.

The 32bit relocs were left alone to follow the same path as others and get assigned 0x0 + addend. Seems like mostly for "if it's not broken" reasons.
Although for non-debug non-alloc can have pointer subtraction:
https://lists.llvm.org/pipermail/llvm-dev/2020-May/141918.html

Anyway. IF my understanding is correct I don't see why we can't generalize for all debug sections to
1) Tombstone to 1 for ranges/loc
2) MAXINT for .debug_names
3) 0 for everything else.

This will leave current behavior for cases that really matter, GC/ICF, and for relative relocations (32 bit outside of .debug_names), I don't think it should matter?

@MaskRay WDYT?

https://github.com/llvm/llvm-project/pull/70701