[llvm-dev] (RFC) Encoding code duplication factor in discriminator

Greg Bedwell via llvm-dev llvm-dev at lists.llvm.org
Tue Nov 22 15:50:36 PST 2016


(resending with history snipped to prevent needing list moderation to post
due to message exceeding size limit)

For our users we fully expect their profile/debug builds to be the same
thing (that is, almost all debugging is on fully optimized code, where
fully optimized also means part of an iterative profiling workflow).
Having extra debug information that only benefits the profiling part seems
reasonable.  What would be unacceptable is if we ever had to ask our users
to make a trade off between debug info that was best for profiling at the
expense of debugging quality or vice versa.  The "-femit-debug-for-profiling"
flag seems to fit this model of it being more a strict superset rather than
a tuning option more clearly.

-Greg

On 22 November 2016 at 23:16, Dehao Chen via llvm-dev <
llvm-dev at lists.llvm.org> wrote:

> -gprofile should have:
>
> 1. emit linkage name in all subprograms
> 2. emit discriminator with dup-factor copy-factor encoded
> 3. use-unknown-locations=true
>
> Here comes the problem: is it an intermediate -g mode between -g1 and -g2?
> If yes, this means that -g2 will also need to include all the 3 items
> above. If not, then it should not be called -gprofile, but rather a flag
> like -femit-debug-for-profiling", so that we can use it with "-g2
> -femit-debug-for-profiling", or "-gmlt -femit-debug-for-profiling".
>
> Dehao
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161122/bbfb5954/attachment.html>


More information about the llvm-dev mailing list