This is an archive of the discontinued LLVM Phabricator instance.

[ELF] - Don't crash when --emit-relocs is used with --gc-sections
ClosedPublic

Authored by grimar on Sep 7 2017, 5:00 AM.

Diff Detail

Repository
rL LLVM

Event Timeline

grimar created this revision.Sep 7 2017, 5:00 AM
ruiu edited edge metadata.Sep 7 2017, 11:28 AM

I don't think this patch addresses the root cause of the issue. All sections passed to addInputSec should be live. If it should be considered dead by gc, it should be marked as such before reaching here. This patch seems to hide the issue instead of fixing it.

phosek added a subscriber: phosek.Sep 7 2017, 3:56 PM
grimar added a comment.EditedSep 8 2017, 12:51 AM
In D37561#863647, @ruiu wrote:

I don't think this patch addresses the root cause of the issue. All sections passed to addInputSec should be live. If it should be considered dead by gc, it should be marked as such before reaching here. This patch seems to hide the issue instead of fixing it.

If that should be true, then why first lines of 'addInputSec' returns (and report discarded sections actually) when sections are not live ?

void OutputSectionFactory::addInputSec(InputSectionBase *IS,
                                       StringRef OutsecName,
                                       OutputSection *&Sec) {
  if (!IS->Live) {
    reportDiscarded(IS);
    return;
  }

What about D37735 ? It should allow to move reportDiscarded out or at least
calls it earlier, what should help this patch hopefully.

ruiu added a comment.Sep 12 2017, 5:32 PM

I think instead of "fixing" a concrete bug, you want to think more abstractly and focus on an invariant at each stage of linking. Invariants are abstract concepts but important to make program easier to understand. One invariant in lld is that, if some section is not marked as "live" after GC, it should be actually live, and it should be emitted to the output file.

A situation like "this section is not marked as live, but this is actually dead because a section that the section depends on is marked as dead by gc" is not ok. That is a violation of the invariant and needs fixing.

grimar planned changes to this revision.Sep 13 2017, 3:10 AM
grimar updated this revision to Diff 115205.EditedSep 14 2017, 5:52 AM
  • Addressed review comment.

Note: D37735 should be landed first for allowing this to work properly. Because at next line:
https://github.com/llvm-mirror/lld/blob/master/ELF/OutputSections.cpp#L230
code would try to access to null pointer Out, when it should just report discarded end return at the begining of addSection method.

ruiu added inline comments.Sep 14 2017, 11:13 AM
ELF/InputSection.cpp
86–88 ↗(On Diff #115205)

This is too magical that it needs comment.

grimar updated this revision to Diff 115382.Sep 15 2017, 2:29 AM
  • Addressed review comment.
ELF/InputSection.cpp
86–88 ↗(On Diff #115205)

Oh, I am sorry, looks I put = instead of |= at last minute by mistake.
Fixed, added comment.

ruiu added inline comments.Sep 15 2017, 2:01 PM
ELF/InputSection.cpp
86–88 ↗(On Diff #115205)

It is better to define a helper function. You can use this.

// Return true if a section with given section flags is live (will never be
// GCed) by default. If a section can be GCed, this function returns false.
static bool isLiveByDefault(uint64_t Flags, uint32_t Type) {
  // If GC is enabled, all memory-mapped sections are subject of GC.
  if (!Config->GcSections)
    return true;
  if (Flags & SHF_ALLOC)
    return false;

  // Besides that, relocation sections can also be GCed because their
  // relocation target sections may be GCed. This doesn't really matter
  // in most cases because the linker usually consumes relocation
  // sections instead of emitting them, but -emit-reloc needs this.
  return Type != SHT_REL && Type != SHT_RELA;
}
grimar updated this revision to Diff 115613.Sep 18 2017, 2:40 AM
  • Addressed review comments.
ELF/InputSection.cpp
86–88 ↗(On Diff #115205)

Used. Thanks !

ruiu accepted this revision.Sep 18 2017, 12:29 PM

LGTM

This revision is now accepted and ready to land.Sep 18 2017, 12:29 PM
This revision was automatically updated to reflect the committed changes.