[Bug tree-optimization/122976] [13/14/15/16 Regression] Miscompile with mixed lto/non-lto, -O3 - stack variable getting clobbered with invalid values

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122976

--- Comment #12 from Martin Jambor <jamborm at gcc dot gnu.org> ---
(In reply to Jacek Sieka from comment #11)
> Thanks for looking into this. Based on the fix, can you think of a
> (reasonable) workaround that we can introduce to make things work with the
> versions of GCC that exhibit the bug? 
> 

As my small testcase shows, the bug is in tree-sra and there is no way
how to 100% avoid it other than disabling the pass altogether with
-fno-tree-sra.  This can however be easily bad for performance.

On the other hand, in your particular testcase, I found out that the
reason why the variable nextRef_1 which is of type
tyObject_Rlp__dNpAlq42D2iOFkwuW8WMyQ is accessed via type
tyTuple__hnCuDR9cLe2lA3YIkwKIyuA (which is the thing that confuses the
pass) is because IPA-ICF (Identical code folding) merges functions
_ZN4sets11eqwasMoved_E3varI5tupleI3int3seqI5uInt8EEE and
_ZN3rlp11eqwasMoved_E3varIN3rlp3RlpEE and as far I can tell(!), such
access is not present in the source.

So you should be able to avoid the issue by compiling at least
@psets.nim.c.i or @m..@srlp.nim.c.i with -fno-ipa-icf.  That can lead
to some code-size growth but hopefully nothing dramatic.