https://gcc.gnu.org/bugzilla/show_bug.cgi?id=27855
--- Comment #56 from Aldy Hernandez <aldyh at gcc dot gnu.org> --- (In reply to Richard Biener from comment #55) > Note the original report used -O and Aldhy used -O2 but we are talking > about a benchmark and when you use -ffast-math you also use -O3. Thanks, I will keep a note of this for next time. FWIW, I used the same flags as the last time this was reconfirmed, which was comment #52. I also saw a bunch of runs at -O2 by Uros, which likely influenced me. > The benchmark is somewhat badly written (manually "optimized") so our > vectorization attempts fail. > > Overall conclusion is I'm unsure if it's worth pursuing this bug further? > There is a register pressure issue left but the testcase maybe not > real-world enough? That is, I would usually recommend to first un-obfuscate > the manually optimized code. IMO, "manually optimized" doesn't seem to merit spending more time on this. As you mention, not real world enough :). May I suggest WONTFIX, but willing to reopen if someone cares enough to un-obfuscate the code?
