[llvm-dev] [SROA][DebugInfo][GSoC] Testing SROA on amalgamated sqlite source

[Adrian Prantl via llvm-dev]

> On May 30, 2018, at 8:37 AM, Anast Gramm <anastasis.gramm2 at gmail.com> wrote:
> 
> Introduction
> ============
> 
>  `SROA' is an early stage pass running at the very beginning of the
>  pipeline in `-O{1,2,3}'. Greg Bedwell's report from his DExTer tool
>  shows SROA on function as one of the major culprits of Debug Info
>  loss.
> 
>  With debugify-each partially done I tried testing this on the
>  amalgamated sqlite source.
> 
>  The steps are as follows:
>  ,----
>  | # generate IR file without debug info and optnone
>  | clang -O0 -Xclang -disable-O0-optnone -S -emit-llvm sqlite3.c -o sqlite
>  | 
>  | # run opt
>  | opt -sroa -debugify-each sqlite -disable-output 2> results
>  `----
> 
> 
> Results
> =======
> 
>  [Here] is the full results file.
> 
>  The results were gather by `grep -c`'ing the results file according to
>  the error's message.
> 
>   SROA runs | FAIL | PASS | empty DebugLoc | Missing line warnings 
>  -----------|------|------|----------------|-----------------------
>        1978 | 1100 |  878 |           5214 |                 75166 
> 
>  What's worth noticing is that the only error that SROA produces is
>  that of empty debug locations. It does not report any missing
>  DILocalVariables or llvm.dbg.value intrinsics. Thus all of the 1100
>  SROA runs that failed are due to empty debug locations.

> 
>  Also, while skimming through the results file I noticed that the
>  instructions with missing DL were `phi' instructions. To confirm this
>  I did:
> 
>    $ grep "ERROR: Instruction" results | grep phi -c
> 
>  witch gives us a result of `5214' and this means that all the
>  instructions with missing Debug Loc are `phi' instructions.
> 
> 
> [Here] https://gramanas.github.io/sroa-results.csv
> 
> 
> Conclusion
> ==========
> 
>  This implies SROA does a pretty good job of preserving debug values.
> 

That is good to hear / expected, since we went though great lengths to ensure that SROA preserves variable debug info. 

Have you looked at the location-less phi instructions? Does empty mean line 0 or do you mean they have no debug location whatsoever? Are they empty for good reasons or does it look more like an oversight in the implementation?

-- adrian

>  The DExTer results are not in line with what I've found though. This
>  might be for a number of reasons. E.g. it could score so low due to
>  the DebugLoc loss. Also debugify doesn't test the backend so it's
>  going to detect a narrower set of issues.
> 
>  Do you have any ideas as to why this happens? Can you confirm the
>  results?
> 
>  @Adrian, @Greg any thoughts would be appreciated!