Enabling debug entry value production by default

Hi all,

I think we’ve reached a state where we’re ready to enable debug entry value production by default for the x86_64, ARM, and AArch64 targets. For context, this is a debug info feature that allows debuggers to recover the value of unmodified optimized-out parameters by ‘going up’ a stack frame and interpreting spilled values, constants, etc. to work out what was passed to the callee. There’s a nice write-up about the impact of this feature in [3]. A big thank you to all of contributors who’ve helped us get here (I’ve tried to CC everyone, sincere apologies if I’ve accidentally left you out).

I believe Djordje and I have shown that there isn’t an outsized binary size or compile time impact of enabling this feature. The aggregate size of DWARF in a stage2 x86_64 RelWithDebInfo clang build grows by under 0.2%, and we didn’t find a significant compile time impact. We’ve shared CTMark results (both compile-time & DWARF section size numbers), as well as a size analysis of stage2 builds of {llc, clang, lldb} in [1].

Djordje has also shared a nice location statistics comparison chart in [1], showing increased variable availability with entry values enabled.

For my part, I instrumented lldb’s DWARFExpression::Evaluate method to log instances in which a variable location is unavailable. I then used a script [4] to single-step through a stage2 RelWithDebInfo+entryvals clang binary as it was compiling sqlite3.c, printing local variables at each stop. Without entry values, lldb reported “variable not available” 883,918 times. With entry values lldb reported “variable not available” 687,899 times (a 22% drop).

Does anyone have concerns about enabling this feature?

thanks,
vedant

[1] [DebugInfo] Enable the debug entry values feature by default
[2] [CallSiteInfo] Fix the assertions regarding updating the CallSiteInfo, [CallSiteInfo] Use the isCandidateForCallSiteEntry() when update the cs info
[3] https://robert.ocallahan.org/2018/11/comparing-quality-of-debug-information.html
[4] https://gist.github.com/vedantk/7602b20a9e1d44c42d32dcca36591cc0

(trying to add back folks dropped by Mail.app…)

  • David Blaikie, David Stenberg, Adrian, Jeremy, and Fred

This would only be enabled when optimizations are enabled, correct?

For GDB and LLDB debug info tuning mode, this sounds good to me. I don’t know if, e.g., the Sony debugger supports entry values. Out of curiosity, does anyone know if GCC enable entry values by default?

– adrian

It has been on by default for -g -O builds in GCC for many years. I see @entry values in backtraces for GCC compiled code with very old GCC 4.7.

Internally in Cisco, we have enabled this feature in LLVM for Cisco code compilation for more than a year now.

-Ananth

Yes, this would just be enabled when optimizations are enabled. Currently, the DWARF emission for entry values is only enabled when tuning for either gdb or lldb.

vedant

For that sort of small growth, if it doesn’t add more relocations (I think the call sites need them (but they’re already emitted/that’s not what we’re discussing enabling here), but call value descriptions don’t), sounds good to me :slight_smile:

Yep, TAG_call_site_parameter and its children shouldn’t require any extra relocations. Thanks!

vedant

I’ve asked someone to check with our (Sony) debugger people, likely hear back early next week. I’d say proceed with having it default on/off based on tuning as you described, and if my guys want to see it turned on, we can fix it later. I have to say, a quick skim through D73534 made me think it’s not controlled that way right now, is that still a TODO?

Thanks for all the work to support this feature!

–paulr

The actual DWARF emission for call site parameters is gated inside of DwarfDebug::constructCallSiteEntryDIEs by `tuneForGDB() || tuneForLLDB()`.

However, we are creating+updating CallSiteInfo (basically, in-memory only bookkeeping used by the backend to keep track of call sites) even when the debugger tuning is set to the Sony debugger. If this creates problems, feel free to file a bug and I'll address it. However, there's an upside to maintaining all this CallSiteInfo even if TAG_call_site_parameters won't be emitted. The backend asserts whenever a MachineInstr carrying CallSiteInfo is destroyed without the call information being updated. This provides useful test coverage.

vedant

This is not how tuning-controlled features are supposed to work. I will comment on the review.

–paulr

Hi,

Thanks you all for the collaboration! :slight_smile:

Paul,

This is not how tuning-controlled features are supposed to work. I will comment on the review.

I see, I am working on addressing the comments from the [1]. I will update the diff asap. Thanks.

Vedant,

There are no entry values generated at -O0 level, but I will add a test case for it. Thanks.

Best regards,
Djordje

Hi,

This has been officially enabled by default.

Thank you all,
Djordje

Thank you, Djordje, Vedant, Paul, for working on it!

A great feature to have in LLVM by default.

Cheers,

Ivan