Debug Info round table notes

IR & Optimizations
,
1

Debug Info round table notes (US LLVM 2024)

Big thanks to @cmtice for taking notes, shared below with a few minor edits, links added etc. And thanks to all attendees for the discussion.

[questions / blanks / post-discussion additions in square brackets]

Interesting DWARF 6 features?

  • Improvements for GPU debugging [proposed?].
  • Prologue start and epilogue end markers to compliment existing prologue end and epilogue begin, to mark regions (supports multiple entry points & more).

Line 0 (compiler generated/unknown/merged loc) regularly an issue for anyone?

  • Sentiment of “yes”, but no metrics or measured impact [graph of unique non-zero locations in CTMark over llvm versions here].
  • Example of recently-ish introduced line 0 bug in Orlando’s talk.
  • Stephen Tozer (Sony) (not present) working on fixing spurious line 0s - fixes coming soon.
  • Steven has a system that attaches stack traces to insns & records where it loses source locations. Auto-regression-detection.
    • (Orlando, Sony) Is it acceptable to automatically notify people if patch drops source locations?
    • (David Blaikie, Google) Should have a human review it & make sure it’s a real issue, then report it to the author. Don’t want to give folks false positives. Be sure to be non-accusatory, polite.
    • Scope for automating once the concept has been proven and used in the field for some time? Possibly.

is_stmt placement (Orlando) See talk / RFC.

RemoveDIs (replace dbg.* intrinsics with DbgRecords) discussion.

  • Timeline:
    • Enabled by default in LLVM 19.
    • Deprecated insertion methods taking instruction pointers (rather than iterators) in LLVM 20.
    • [Aim to remove dbg.* intrinsic support from LLVM 21.]
    • Patches caused anyone issues? (Adrian Prantl, Apple) Some, downstream porting in Swift compiler; not too bad yet.
  • How close are we to “debug info doesn’t affect codegen”?
    • RemoveDIs fixed ~70% of an internal test suite at Sony.
    • RemoveDIs is currently only in LLVM IR - we still have debug instructions in MIR.
    • Instruction Referencing (below) may enable MIR changes.

Instruction Referencing (InstrRef, alt LiveDebugValues impl).

  • SSA based variable location tracking in MIR.
  • Improves variable locations (over old version) but requires target support (only X86 supported currently).
  • Hope to remove old version one day but need to bring up support for other targets.
  • (Shubham Rastogi, Apple) hoping to look at AArch64 support.
  • Turning on for unsupported target
  • Others [?] noted PowerPC [and another?] support is important, but who to do the work?

Variable location coverage metrics

  • Student talk: “A data-driven approach to debug info quality” (Emil Pedersen, Apple). Inserted analysis into Swift compiler to measure how much var loc info got dropped in each compiler pass.
  • -mllvm -dropped-variable-stats, StandardInstrumentations.h.
  • (Shubham) Patch in LLVM to reduce false positives (e.g., after DSE).
  • Prints info about how many variables were dropped in particular passes, to stdout.
  • Can we track these over time? Designed/thought about but not implemented.

DW_AT_trampoline

  • To allow transparently stepping through wrappers/thunks to callee.
  • ([?], Nintendo) current state of patches adding support in LLVM?
  • Augusto Noronha at Apple has patches supporting one [?] attribute value: … [from memory, references?]

DW_OP_implicit_pointers - interest, and anyone working on it?

  • Sony & Fuchsia indicate interest. No one actively working on it yet.
  • Likely measurable fraction of variable location coverage improvement.

Long DIExpression / location expressions

  • (Adrian) noticed v long expressions, hitting internal LLVM cap (=> kills var loc).
  • (Shubham) added basic constant folding, DIExpression::foldConstantMath, applied after salvaging [and in DIExpression::append].

Please feel free to add anything missed, suggest edits, fill in blanks, etc.

5 Likes
LLVM Weekly - #566, November 4th 2024