[Core] Do not reclaim absolute atoms in resolver

[davide]

Summary

After discussion with Rui, he explained me the Resolver shouldn't reclaim absolute atoms at all.
This was found out while testing --gc-sections for the GNU driver, which cause lld to crash because of reclaim of absolute atoms.
The patch proposed is supposed to fix the problem.
I modified an existing test rather than adding a new one. Let me know if it's fine for you.

[shankarke]

lib/Core/Resolver.cpp

393–394

Is this something GNU only ?

[shankarke]

lib/Core/Resolver.cpp

393–394

If its GNU only, then you need to add a linkingContext get/set function to check for flavors that want to reclaim absolute atoms.

[ruiu]

LGTM. I don't believe this is really ELF specific, it's harmless on other platforms at least. Let's just go ahead and see how it goes.

[shankarke]

I feel its a bug in GNU on why absolute symbols are not reclaimed. Do we want to be bug compatible with lld ?

cat > 1.c << \!
int foo() { return 0; }
!

cat > 2.c << \!
int bar() { return 0; }
!

gcc -c 1.c 2.c
ld 1.o 2.o -e foo --gc-sections

nothing in bar is used, but still there is an absolute symbol, confuses me.

[shankarke]

There is a absolute symbol for 2.c symbol.

readelf shows this :-

1: 0000000000000000     0 FILE    LOCAL  DEFAULT  ABS 2.c

[ruiu]

Can you elaborate why you think that's a bug of the GNU linker? Note that
absolute symbols are not sections.

[shankarke]

nothing from 2.o was really needed in the final object. absolute symbols are specific to a file, its easy to enhance GC to collect them IMO. This is also needed by --defsym too, if there is not a reference to that symbol, ld doesnot generate one.

[ruiu]

That we can "extend" the existing behavior of the GNU linker doesn't mean
that's a bug of the GNU linker. I don't even agree that it's a reasonable
extension, as --gc-sections means gc'ing sections and absolute symbols
don't belong to any section, but anyways, if you want to do something
different from the GNU linker, we need a really strong justification for
doing that. Making it compatible is the point of creating a compatible
driver.

lib/Core/Resolver.cpp

 
   // Some type of references prevent referring atoms to be dead-striped.
   // Make a reverse map of such references before traversing the graph.
-  for (const Atom *atom : _atoms)
+  // While traversing the list of atoms, mark AbsoluteAtoms as live
+  // in order to avoid reclaim.
+  for (const Atom *atom : _atoms) {
     if (const DefinedAtom *defAtom = dyn_cast<DefinedAtom>(atom))
       for (const Reference *ref : *defAtom)
         if (isBackref(ref))
           _reverseRef[ref->target()].insert(atom);
+    if (const AbsoluteAtom *absAtom = dyn_cast<AbsoluteAtom>(atom))
+      markLive(absAtom);
+  }
 
   // By default, shared libraries are built with all globals as dead strip roots
   if (_context.globalsAreDeadStripRoots())

test/core/absolute-basic.objtxt

-# RUN: lld -core %s | FileCheck %s
+# RUN: lld -core --dead-strip %s | FileCheck %s
 
 #
 # Test that absolute symbols are parsed and preserved