[fpc-devel] Going to properly investigate that 3.0.2 bug

[Martin Frb]

On 10/07/2021 06:40, J. Gareth Moreton via fpc-devel wrote:
> Hi everyone,
>
> So I want to fix that optimisation bug in FPC 3.0.2 and 3.2.0 etc, but 
> I need a little bit of help.  I know that Lazarus doesn't build 
> properly with optimisations, but does anyone know of a smaller project 
> that compiles incorrectly and causes an easy crash when run?  If not, 
> what do I need to do with Lazarus to make it crash?  With that, I can 
> do a kind of manual binary search with revisions to find the point 
> where it no longer crashes (since it was fixed for FPC 3.3.1).

The one that I got is very hard to catch.

It seem to me, that something is doing a write to random memory. So any 
tiny change, and the bug disappears.


This is the revision that for me reproduces it.
https://github.com/User4martin/lazarus/tree/f-fpc-crash-fpdebug

The testcase is
components\lazdebuggers\lazdebuggerfp\test\LazDebFpTest.lpi
build mode "64bit -O2"

To run it you need to copy    fpclist.txt.sample   to  fpclist.txt
edit it, and point it to your fpc (can be an older, can be trunk, does 
not have to be the crashing version)
you need only one fpc in it.

You also need to create some folders
logs/
lib/
testapps/lib/

And it is Windows 64 bit only.

Make sure it crashes  for you, before trying to hunt down what it is.

-------------------------------------

I tried to create a console test case => but then it does not crash.

If you use OoNOPEEPHOLE  => no crash

If you use "implicit exception off" => no crash
see https://github.com/User4martin/lazarus/tree/f-fpdebug-fpc-optimizer-bug

I did trace it down to the specialization of some generic, that needed 
to be done while "implicit exception off"
But that is a lot of functions...

-----
All that said, I know there is no proof the bug is not in my code.
It's just based on likelyhood.

I valgrinded the crap out of it. Nothing.
Well it is only windows, but if I accessed a dangling pointer, valgrind 
would find it, even if it caused no harm on Linux.
So the bug is apparently not in the compiled code on Linux.

And that is basically all I have....


--------------------------
you can edit the project file and change

   TestRunner.RunExecute(TestRunner);
   //Application.Run;

then the test runs automatically.

Or even
   TestRunner.Show;
TestRunner.RunTest(TTest(TestRunner.TestTree.Items[0].Items[2].Data));

To run just a small subset of tests (you might need to change that 
depending on your fpclist.txt)