[lld] [lld][ELF] Add --why-live flag (inspired by Mach-O) (PR #127112)

[Daniel Thornburgh via llvm-commits]

================
@@ -119,17 +142,29 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
     // group/SHF_LINK_ORDER rules (b) if the associated text section should be
     // discarded, marking the LSDA will unnecessarily retain the text section.
     if (!(fromFDE && ((relSec->flags & (SHF_EXECINSTR | SHF_LINK_ORDER)) ||
-                      relSec->nextInSectionGroup)))
-      enqueue(relSec, offset);
+                      relSec->nextInSectionGroup))) {
+      Symbol *canonicalSym = d;
+      if (TrackWhyLive && d->isSection()) {
+        if (Symbol *s = relSec->getEnclosingSymbol(offset))
----------------
mysterymath wrote:

I think it's possible, but it AFAICT it would be quite complicated.

The other instance of `getEnclosingSymbol` was easy to defer, since it only affected the right hand side of the `whyLive` map. Deferring determination of identity on the left hand side of the map is different. In particular, it wouldn't be clear whether a given `SecOffset` actually does match the `--why-live` glob without resolving it. It also makes it difficult to maintain the "first reference wins" invariant that makes the `whyLive` graph acyclic. Multiple `SecOffset`s could end up resolving to the same symbol, and the resolved reason would need to be the first of all such references in the order of discovery.

I've added some comments to document the invariant and the reasoning for leaving this in. I'd expect that walking the symbols for a referenced section wouldn't be too bad in the usual `--function-sections` `--data-sections` case, but if that's wrong, we could build something like the following:
 - Keep an order-of-discovery counter for each deferred `SecOffset`
 - Walk all symbols and collect sections that contain symbols that match globs and unresolved `SecOffset`s
 - For each such section:
   - Collect the unresolved offsets and sort them
   - Walk the symbols in value order and resolve the offsets in one pass
 - While printing, if a section is encountered that has unresolved `SecOffset`s, do the above routine too.
 
I suppose we could also just throw up our hands and build a symbol interval map for each section when doing whyLive. That would probably be dramatically simpler.

https://github.com/llvm/llvm-project/pull/127112