RE: Eigen 3.3 vs 3.2 Performance (was RE: [eigen] 3.3-beta2 released!)

[ Thread Index | Date Index | More Archives ]

On Tue, 21 Aug 2018, Daniel.Vollmer@xxxxxx wrote:

I'm not quite sure how to report the gcc-8 problem to the gcc-maintainers. I've tried running creduce over the code, but the output (attached) seems slightly different (e.g. the copy in operator+=, which looks like the aliasing assignment was chosen).

I've reported
(the source may be large and unreduced, but internally after the first few rounds of optimization it is rather clear what is going on so it should be good enough for gcc devs)

There were other regressions in gcc-8 related to the same optimization pass (SRA) : 85459, 85762, etc. IIRC the author said he was trying to rewrite that pass for gcc-9, but I could be wrong.

It might not be the only issue with gcc-8 though.

Marc Glisse

Mail converted by MHonArc 2.6.19+