Re: Side effect of -finline-functions

[Alexander Monakov <amonakov@xxxxxxxxx>]

On Thu, 3 May 2012, Mathieu Malaterre wrote:

> 
> Hi all,
> 
>   We are currently tracking an issue within the openjpeg codebase. One
> user reported that compilation with -O3 leads to incorrect result
> decompressing a particular image. He also noticed that compilation
> with -O2 -fgcse-after-reload  -fipa-cp-clone -fpredictive-commoning
> -ftree-vectorize -funswitch-loops does not produce any artefact.
> Therefore we concluded that the only possible side effect would be
> generated by the flag -finline-functions
> 
> Is there some kind of tools that could help us locate exactly which
> function is producing a different output with this flag ?

I'm not aware of such tool.  However, what you can do is to find a file that
contains a function that behaves differently with -finline-functions flag. By
splitting that file further, you may be able to find a set of functions that
behave differently with inlining.

To do that, prepare a two sets of openjpeg's object files: one "good" without
-finline-functions flag used for compilation, and the other "bad" -- with that
flag.  Then compile the executable from complementing subsets of each set,
until a single file from "bad" set is proven to produce an incorrect result.
Refer to http://gcc.gnu.org/wiki/Analysing_Large_Testcases for scripts that
automate this job (ignore the part about debug counters as inlining does not
have any).

Hope that helps,
Alexander