Description
| Bugzilla Link | 23837 |
| Version | trunk |
| OS | Linux |
| Reporter | LLVM Bugzilla Contributor |
| CC | @dwblaikie,@echristo,@eugenis,@pogo59 |
Extended Description
Generally, LLVM does very bad job in ensuring that, when debug info is enabled (-gline-tables-only or above), all necessary([1]) instructions indeed have DebugLoc attached to them, or that DebugLoc are preserved as we transform the code. For example:
- DebugLoc can be easily missed as we split/rearrange basic blocks.
- Synthesized instructions in GVN, SCEV etc. often have no debug info.
Once we have an instruction I with no DebugLoc, this "missing" DebugLoc would propagate through the program - e.g. every time we construct IRBuilder<> with I as insertion point, it will fail to set correct DebugLoc for instructions it creates.
This is painful - with optimizations enabled, we may often fail to provide reasonable line info for instructions, or (worse case, which often happens) incorrectly return line info for some preceding instruction which happened
to have DebugLoc, but may be completely irrelevant (e.g. come from a
different inlined subroutine).
I've did several improvements recently:
http://llvm.org/viewvc/llvm-project?rev=239438&view=rev
http://llvm.org/viewvc/llvm-project?rev=239479&view=rev
http://llvm.org/viewvc/llvm-project?rev=239550&view=rev
http://llvm.org/viewvc/llvm-project?rev=239551&view=rev
http://llvm.org/viewvc/llvm-project?rev=239584&view=rev
http://llvm.org/viewvc/llvm-project?rev=239585&view=rev
http://llvm.org/viewvc/llvm-project?rev=239643&view=rev
But there are many more cases to fix, and even if we fix them all,
we should be able to somehow prevent further breakages.
- We can improve API for creating new instructions that would force callers
to set DebugLoc (or explicitly pass empty one, promising to set it up later). - Add module verifier check that would enforce that if function has some
attached DebugLocs, then all ([1]) instruction should have them. - Add asserts to places where we believe DebugLoc should always be present if
line tables are enabled (e.g. we need it in SanitizerCoverage pass where
we know that PCs of certain instructions would be collected, and possibly
symbolized later). - Some combination of above.
Notes:
[1] - Function prologue instructions (alloca's etc.) intentionally have no DebugLoc. PHI nodes sometimes do have them, but it's not clear whether there's a contract of some sort. Many users actually call BB->getFirstNonPHI()->getDebugLoc() to get the "entry" debug location, probably because they believe it will be the first insertion point / non-transient instruction.