[Bug target/84616] funsafe-math-optimizations leads to incorrect results for 4x4 matrix inversion

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84616

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |needs-reduction, wrong-code
             Target|                            |x86_64-*-*, i?86-*-*
             Status|WAITING                     |NEW
                 CC|                            |rguenth at gcc dot gnu.org
          Component|c++                         |target
            Version|unknown                     |7.3.1

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Hmm, I get (with local Eigen library) at -O0:

 1 -0  0 -0
-0  1 -0  0
 0 -0  1 -0
-0  0 -0  1

so it looks like Eigen has issues with signed zeros?  I can reproduce your
bogus output already with -O -fno-signed-zeros with my eigen version.

Your preprocessed source also produces those signed zero issues seen above
but requires -funsafe-math-optimizations to reproduce the issue.

So, "confirmed", but given the complexity of Eigen I wouldn't rule out
issues with the library itself.

Interestingly with -m32 -Ofast -mno-sse I see correct results (apart from
the signed zeros).

Needs to be tracked down to sth simpler than Eigen...